Image Biomarker Standardisation Initiative Reference Manual Arxiv:1612.07003V11 [Cs.CV] 17 Dec 2019 the Image Biomarker Standardisation Initiative

Image biomarker standardisation initiative Reference manual arXiv:1612.07003v11 [cs.CV] 17 Dec 2019 The image biomarker standardisation initiative

The image biomarker standardisation initiative (IBSI) is an independent international collaboration which works towards standardising the extraction of image biomarkers from acquired imaging for the purpose of high-throughput quantitative image analysis (radiomics). Lack of reproducibility and validation of radiomic studies is considered to be a major challenge for the field. Part of this challenge lies in the scantiness of consensus-based guidelines and definitions for the process of translating acquired imaging into high-throughput image biomarkers. The IBSI therefore seeks to provide standardised image biomarker nomenclature and definitions, a standardised general image processing workflow, tools for verifying radiomics software implementations and reporting guidelines for radiomic studies.

Permanent identifiers The IBSI uses permanent identifiers for image biomarker definitions and important related concepts such as image processing. These consist of four-character codes and may be used for reference. Please do not use page numbers or section numbers as references, as these are subject to change.

Copyright This work is a copy-edited version of the final (v10) pre-print version of the IBSI reference manual, which was licensed under the Creative Commons Attribution 4.0 International License (CC-BY). The original work may be cited as: Zwanenburg A, Leger S, VallièresM, Löck S. Image biomarker standardisation initiative. arXiv preprint arXiv:1612.07003. This document is licensed under the Creative Commons Attribution 4.0 International License. To view a copy of this license, visit http://creativecommons.org/licenses/by/4.0/ or send a letter to Creative Commons, PO Box 1866, Mountain View, CA 94042, USA. The digital phantom (see section 5.1) is licensed under the Creative Commons Attribution 4.0 International License. To view a copy of this license, visit http://creativecommons.org/licenses/by/4.0/ or send a letter to Creative Commons, PO Box 1866, Mountain View, CA 94042, USA. The radiomics phantom (see section 5.2), which is based on a human lung cancer computed tomography image and published by cancerdata.org (DOI:10.17195/candat.2016.08.1), is licensed under the Creative Commons Attribution-NonCommercial 3.0 Unported Licence. To view a copy of this license, visit https://creativecommons.org/licenses/by-nc/3.0/ or send a letter to Creative Commons, PO Box 1866, Mount View, CA 94042, USA. This license pertains to both the original DICOM set, as well as the same data in NIfTI format released by the IBSI. Copyright information regarding the reference data sets may be found on GitHub: https://github.com/theibsi/data sets

i ii

Citation information To cite the document or the digital phantom, please use the following citation:

1. Zwanenburg A, Leger S, Valli`eresM, L¨ock S. Image biomarker standardisation initiative. arXiv preprint arXiv:1612.07003.

Additionally, when using the radiomics phantom originally published on cancerdata.org, please include the following citation:

1. Lambin P, Leijenaar RT, Deist TM, Peerlings J, de Jong EE, van Timmeren J, Sanduleanu S, Larue RT, Even AJ, Jochems A, van Wijk Y. Radiomics: the bridge between medical imaging and personalized medicine. Nature Reviews Clinical Oncology. 2017 Dec;14(12):749.

2. Lambin P. Radiomics Digital Phantom, CancerData (2016), DOI:10.17195/candat.2016.08.1

Contact Dr. Alex Zwanenburg [email protected] Dr. Martin Valli`eres [email protected]

IBSI contributors

Mahmoud A. Abdalah Image Response Assessment Team Core Facility, Moﬃtt Cancer Center, Tampa (FL), USA Hugo J.W.L. Aerts Dana-Farber Cancer Institute, Brigham and Women’s Hospital, and Harvard Medical School, Harvard University, Boston (MA), USA Vincent Andrearczyk Institute of Information Systems, University of Applied Sciences Western Switzerland (HES-SO), Sierre, Switzerland Aditya Apte Department of Medical Physics, Memorial Sloan Kettering Cancer Center, New York (NY), USA Saeed Ashraﬁnia Department of Electrical and Computer Engineering, Johns Hop- kins University, Baltimore (MD), USA and Department of Radi- ology and Radiological Science, Johns Hopkins University, Bal- timore (MD), USA Spyridon Bakas Center for Biomedical image Computing and Analytics (CBICA), University of Pennsylvania, Philadelphia (PA), USA and Depart- ment of Radiology, Perelman School of Medicine, University of Pennsylvania, Philadelphia (PA), USA and Department of Patho- logy and Laboratory Medicine, Perelman School of Medicine, Uni- versity of Pennsylvania, Philadelphia (PA), USA Roelof J. Beukinga Department of Nuclear Medicine and Molecular Imaging, Univer- sity of Groningen, University Medical Center Groningen (UMCG), Groningen, The Netherlands Ronald Boellaard Department of Nuclear Medicine and Molecular Imaging, Univer- sity of Groningen, University Medical Center Groningen (UMCG), Groningen, The Netherlands and Radiology and Nuclear Medi- cine, VU University Medical Centre (VUMC), Amsterdam, The Netherlands continued on next page iii

Marta Bogowicz Department of Radiation Oncology, University Hospital Zurich, University of Zurich, Zurich, Switzerland Luca Boldrini Fondazione Policlinico Universitario ””A. Gemelli”” IRCCS, Roma, Italia IrèneBuvat Imagerie Moléculaire In Vivo, CEA, Inserm, Univ Paris Sud, CNRS, Universit Paris Saclay, Orsay, France Gary J.R. Cook Cancer Imaging Dept, School of Biomedical Engineering and Ima- ging Sciences, Kings College London, London, United Kingdom Christos Davatzikos Center for Biomedical image Computing and Analytics (CBICA), University of Pennsylvania, Philadelphia (PA), USA and Depart- ment of Radiology, Perelman School of Medicine, University of Pennsylvania, Philadelphia (PA), USA Adrien Depeursinge Institute of Information Systems, University of Applied Sciences Western Switzerland (HES-SO), Sierre, Switzerland and Depart- ment of Nuclear Medicine and Molecular Imaging, Lausanne Uni- versity Hospital, Lausanne, Switzerland Marie-Charlotte Desseroit Laboratory of Medical Information Processing (LaTIM) - team ACTION (image-guided therapeutic action in oncology), IN- SERM, UMR 1101, IBSAM, UBO, UBL, Brest, France Nicola Dinapoli Fondazione Policlinico Universitario ””A. Gemelli”” IRCCS, Roma, Italia Cuong Viet Dinh Department of Radiation Oncology, the Netherlands Cancer In- stitute (NKI), Amsterdam, The Netherlands Sebastian Echegaray Department of Radiology, Stanford University School of Medicine, Stanford (CA), USA Issam El Naqa Department of Radiation Oncology, Physics Division, University of Michigan, Ann Arbor (MI), USA and Medical Physics Unit, McGill University, Montréal,Qubec, Canada Andriy Y. Fedorov Surgical Planning Laboratory, Brigham and Women’s Hospital and Harvard Medical School, Harvard University, Boston (MA), USA Roberto Gatta Fondazione Policlinico Universitario ””A. Gemelli”” IRCCS, Rome, Italy Robert J. Gillies Department of Cancer Imaging and Metabolism, Moffitt Cancer Center, Tampa (FL), USA Vicky Goh Cancer Imaging Dept, School of Biomedical Engineering and Ima- ging Sciences, Kings College London, London, United Kingdom Matthias Guckenberger Department of Radiation Oncology, University Hospital Zurich, University of Zurich, Zurich, Switzerland Michael Götz Department of Medical Image Computing, German Cancer Re- search Center (DKFZ), Heidelberg, Germany Sung Min Ha Center for Biomedical image Computing and Analytics (CBICA), University of Pennsylvania, Philadelphia (PA), USA and Depart- ment of Radiology, Perelman School of Medicine, University of Pennsylvania, Philadelphia (PA), USA Mathieu Hatt Laboratory of Medical Information Processing (LaTIM) - team ACTION (image-guided therapeutic action in oncology), IN- SERM, UMR 1101, IBSAM, UBO, UBL, Brest, France continued on next page iv

Fabian Isensee Department of Medical Image Computing, German Cancer Re- search Center (DKFZ), Heidelberg, Germany Philippe Lambin The D-Lab, Department of Precision Medicine, GROW-School for Oncology and Developmental Biology, Maastricht University Medical Centre+, Maastricht, The Netherlands Baptiste Laurent Laboratory of Medical Information Processing (LaTIM) - team ACTION (image-guided therapeutic action in oncology), IN- SERM, UMR 1101, IBSAM, UBO, UBL, Brest, France Stefan Leger National Center for Tumor Diseases (NCT), Partner Site Dresden, Germany: German Cancer Research Center (DKFZ), Heidel- berg, Germany; Faculty of Medicine and University Hospital Carl Gustav Carus, Technische UniversitätDresden, Dresden, Ger- many, and; Helmholtz Association / Helmholtz-Zentrum Dresden - Rossendorf (HZDR), Dresden, Germany and OncoRay – Na- tional Center for Radiation Research in Oncology, Faculty of Medicine and University Hospital Carl Gustav Carus, Technische Universitt Dresden, Helmholtz-Zentrum Dresden - Rossendorf, Dresden, Germany and German Cancer Consortium (DKTK), Partner Site Dresden, and German Cancer Research Center (DKFZ), Heidelberg, Germany Ralph T.H. Leijenaar The D-Lab, Department of Precision Medicine, GROW-School for Oncology and Developmental Biology, Maastricht University Medical Centre+, Maastricht, The Netherlands Jacopo Lenkowicz Fondazione Policlinico Universitario ””A. Gemelli”” IRCCS, Rome, Italy Fiona Lippert Section for Biomedical Physics, Department of Radiation Onco- logy, University of Tübingen,Germany Are Losneg˚ard Department of Clinical Medicine, University of Bergen, Bergen, Norway Steffen Löck OncoRay – National Center for Radiation Research in Oncology, Faculty of Medicine and University Hospital Carl Gustav Carus, Technische UniversitätDresden, Helmholtz-Zentrum Dresden - Rossendorf, Dresden, Germany and German Cancer Consortium (DKTK), Partner Site Dresden, and German Cancer Research Center (DKFZ), Heidelberg, Germany and Department of Radio- therapy and Radiation Oncology, Faculty of Medicine and Univer- sity Hospital Carl Gustav Carus, Technische Universitt Dresden, Dresden, Germany Klaus H. Maier-Hein Department of Medical Image Computing, German Cancer Re- search Center (DKFZ), Heidelberg, Germany Sarah A. Mattonen Department of Medical Biophysics, The University of Western Ontario, London (ON), Canada Olivier Morin Department of Radiation Oncology, University of California, San Francisco (CA), USA Henning Müller Institute of Information Systems, University of Applied Sciences Western Switzerland (HES-SO), Sierre, Switzerland and Univer- sity of Geneva, Geneva, Switzerland continued on next page v

Sandy Napel Department of Radiology, Stanford University School of Medi- cine, Stanford (CA), USA and Department of Electrical Engin- eering, Stanford University, Stanford (CA), USA and Department of Medicine (Biomedical Informatics Research), Stanford Univer- sity School of Medicine, Stanford (CA), USA Christophe Nioche Imagerie Moléculaire In Vivo, CEA, Inserm, Univ Paris Sud, CNRS, Universit Paris Saclay, Orsay, France Fanny Orlhac Imagerie Moléculaire In Vivo, CEA, Inserm, Univ Paris Sud, CNRS, Universit Paris Saclay, Orsay, France Sarthak Pati Center for Biomedical image Computing and Analytics (CBICA), University of Pennsylvania, Philadelphia (PA), USA and Depart- ment of Radiology, Perelman School of Medicine, University of Pennsylvania, Philadelphia (PA), USA Elisabeth A.G. Pfaehler Department of Nuclear Medicine and Molecular Imaging, Univer- sity of Groningen, University Medical Center Groningen (UMCG), Groningen, The Netherlands Arman Rahmim Departments of Radiology and Physics, University of British Columbia, Vancouver (BC), Canada and Department of Radi- ology and Radiological Science, Johns Hopkins University, Bal- timore (MD), USA Arvind U.K. Rao Department of Computational Medicine and Bioinformatics, Uni- versity of Michigan, Ann Arbor (MI), USA and Department of Ra- diation Oncology, Physics Division, University of Michigan, Ann Arbor (MI), USA Christian Richter OncoRay – National Center for Radiation Research in Oncology, Faculty of Medicine and University Hospital Carl Gustav Carus, Technische UniversitätDresden, Helmholtz-Zentrum Dresden - Rossendorf, Dresden, Germany and German Cancer Consortium (DKTK), Partner Site Dresden, and German Cancer Research Center (DKFZ), Heidelberg, Germany and Department of Radio- therapy and Radiation Oncology, Faculty of Medicine and Univer- sity Hospital Carl Gustav Carus, Technische Universitt Dresden, Dresden, Germany and Helmholtz-Zentrum Dresden - Rossendorf, Institute of Radiooncology OncoRay, Dresden, Germany Jonas Scherer Department of Medical Image Computing, German Cancer Re- search Center (DKFZ), Heidelberg, Germany Muhammad Musib Siddique Cancer Imaging Dept, School of Biomedical Engineering and Ima- ging Sciences, Kings College London, London, United Kingdom Nanna M. Sijtsema Department of Radiation Oncology, University of Groningen, University Medical Center Groningen (UMCG), Groningen, The Netherlands Jairo Socarras Fernandez Section for Biomedical Physics, Department of Radiation Onco- logy, University of Tübingen,Germany Emiliano Spezi School of Engineering, Cardiff University, Cardiff, United King- dom and Department of Medical Physics, Velindre Cancer Centre, Cardiff, United Kingdom Roel J.H.M Steenbakkers Department of Radiation Oncology, University of Groningen, University Medical Center Groningen (UMCG), Groningen, The Netherlands continued on next page vi

Stephanie Tanadini-Lang Department of Radiation Oncology, University Hospital Zurich, University of Zurich, Zurich, Switzerland Daniela Thorwarth Section for Biomedical Physics, Department of Radiation Onco- logy, University of Tübingen,Germany Esther G.C. Troost OncoRay – National Center for Radiation Research in Oncology, Faculty of Medicine and University Hospital Carl Gustav Carus, Technische UniversitätDresden, Helmholtz-Zentrum Dresden - Rossendorf, Dresden, Germany and National Center for Tumor Diseases (NCT), Partner Site Dresden, Germany: German Cancer Research Center (DKFZ), Heidelberg, Germany; Faculty of Medi- cine and University Hospital Carl Gustav Carus, Technische Uni- versitt Dresden, Dresden, Germany, and; Helmholtz Association / Helmholtz-Zentrum Dresden - Rossendorf (HZDR), Dresden, Germany and German Cancer Consortium (DKTK), Partner Site Dresden, and German Cancer Research Center (DKFZ), Heidelberg, Germany and Department of Radiotherapy and Radi- ation Oncology, Faculty of Medicine and University Hospital Carl Gustav Carus, Technische Universitt Dresden, Dresden, Germany and Helmholtz-Zentrum Dresden - Rossendorf, Institute of Ra- diooncology OncoRay, Dresden, Germany Taman Upadhaya Department of Nuclear Medicine, CHU Milétrie,Poitiers, France and Laboratory of Medical Information Processing (LaTIM) - team ACTION (image-guided therapeutic action in oncology), IN- SERM, UMR 1101, IBSAM, UBO, UBL, Brest, France Vincenzo Valentini Fondazione Policlinico Universitario ””A. Gemelli”” IRCCS, Rome, Italy Martin Vallières Medical Physics Unit, McGill University, Montréal, Qubec, Canada Lisanne V. van Dijk Department of Radiation Oncology, University of Groningen, University Medical Center Groningen (UMCG), Groningen, The Netherlands Joost van Griethuysen Department of Radiology, the Netherlands Cancer Institute (NKI), Amsterdam, The Netherlands and GROW-School for On- cology and Developmental Biology, Maastricht University Medical Center, Maastricht, The Netherlands and Department of Radi- ation Oncology, Dana-Farber Cancer Institute, Brigham and Wo- mens Hospital, Harvard Medical School, Boston (MA), USA Floris H.P. van Velden Department of Radiology, Leiden University Medical Center (LUMC), Leiden, The Netherlands Philip Whybra School of Engineering, Cardiff University, Cardiff, United King- dom continued on next page vii

Alex Zwanenburg OncoRay – National Center for Radiation Research in Oncology, Faculty of Medicine and University Hospital Carl Gustav Carus, Technische Universit¨atDresden, Helmholtz-Zentrum Dresden - Rossendorf, Dresden, Germany and National Center for Tumor Diseases (NCT), Partner Site Dresden, Germany: German Cancer Research Center (DKFZ), Heidelberg, Germany; Faculty of Medi- cine and University Hospital Carl Gustav Carus, Technische Uni- versitt Dresden, Dresden, Germany, and; Helmholtz Association / Helmholtz-Zentrum Dresden - Rossendorf (HZDR), Dresden, Germany and German Cancer Consortium (DKTK), Partner Site Dresden, and German Cancer Research Center (DKFZ), Heidel- berg, Germany

Table 1 — Alphabetical list of IBSI contributors. Contents

1 Introduction 1

2 Image processing 3 2.1 Data conversion ...... 3 2.2 Post-acquisition processing ...... 5 2.3 Segmentation ...... 5 2.4 Interpolation ...... 6 2.5 Re-segmentation ...... 11 2.6 ROI extraction ...... 13 2.7 Intensity discretisation ...... 13 2.8 Feature calculation ...... 15

3 Image features 17 3.1 Morphological features ...... 20 3.2 Local intensity features ...... 35 3.3 Intensity-based statistical features ...... 37 3.4 Intensity histogram features ...... 45 3.5 Intensity-volume histogram features ...... 55 3.6 Grey level co-occurrence based features ...... 59 3.7 Grey level run length based features ...... 82 3.8 Grey level size zone based features ...... 96 3.9 Grey level distance zone based features ...... 106 3.10 Neighbourhood grey tone diﬀerence based features ...... 116 3.11 Neighbouring grey level dependence based features ...... 122

4 Radiomics reporting guidelines and nomenclature 133 4.1 Reporting guidelines ...... 133 4.2 Feature nomenclature ...... 138

5 Reference data sets 144

viii CONTENTS ix

5.1 Digital phantom ...... 144 5.2 Lung cancer CT image ...... 145

A Digital phantom texture matrices 150 A.1 Grey level co-occurrence matrix (2D) ...... 150 A.2 Grey level co-occurrence matrix (2D, merged) ...... 151 A.3 Grey level co-occurrence matrix (3D) ...... 152 A.4 Grey level co-occurrence matrix (3D, merged) ...... 154 A.5 Grey level run length matrix (2D) ...... 154 A.6 Grey level run length matrix (2D, merged) ...... 156 A.7 Grey level run length matrix (3D) ...... 156 A.8 Grey level run length matrix (3D, merged) ...... 158 A.9 Grey level size zone matrix (2D) ...... 158 A.10 Grey level size zone matrix (3D) ...... 158 A.11 Grey level distance zone matrix (2D) ...... 159 A.12 Grey level distance zone matrix (3D) ...... 159 A.13 Neighbourhood grey tone diﬀerence matrix (2D) ...... 159 A.14 Neighbourhood grey tone diﬀerence matrix (3D) ...... 159 A.15 Neighbouring grey level dependence matrix (2D) ...... 160 A.16 Neighbouring grey level dependence matrix (3D) ...... 160 Chapter 1

Introduction

A biomarker is ”a characteristic that is objectively measured and evaluated as an indicator of normal biological processes, pathogenic processes, or pharmacologic responses to a therapeutic in- tervention” 7. Biomarkers may be measured from a wide variety of sources, such as tissue samples, cell plating, and imaging. The latter are often referred to as imaging biomarkers 55. Imaging biomarkers consist of both qualitative biomarkers, which require expert interpretation, and quantitative biomarkers which are based on mathematical definitions. Calculation of quantitative imaging biomarkers can be automated, which enables high-throughput analyses. We refer to such (high- throughput) quantitative biomarkers as image biomarkers to differentiate them from qualitative imaging biomarkers. Image biomarkers characterise the contents of (regions of) an image, such as volume or mean intensity. Because of the historically close relationship with the computer vision field, image biomarkers are also referred to as image features. The term features, instead of biomarkers, will be used throughout the remainder of the reference manual, as the contents are generally applicable and not limited to life sciences and medicine only. This work focuses specifically on the (high-throughput) extraction of image biomarkers from acquired, reconstructed and stored imaging. High-throughput quantitative image analysis (radiomics) has shown considerable growth in e.g. cancer research 41, but the scarceness of consensus guidelines and definitions has led to it being described as a ”wild frontier” 13. This reference manual therefore presents an effort to chart a course through part of this frontier by presenting consensus-based recommendations, guidelines, definitions and reference values for image biomarkers and defining a general radiomics image processing scheme. We hope use of this manual will improve reproducibility of radiomic studies. We opted for a specific focus on the computation of image biomarkers from acquired imaging. Thus, validation of imaging biomarkers, either viewed in a broader framework such as the one presented by O’Connor et al. 55 , or within smaller-scope settings such as those presented by Caicedo et al. 13 and by Lambin et al. 41 , falls beyond the scope of this work. Notably, the issue of har- monising and standardising (medical) image acquisition and reconstruction is being addressed in a more comprehensive manner by groups such as the Quantitative Imaging Biomarker Alliance 53,68, the Quantitative Imaging Network 17,54, and task groups and committees of the American Associ- ation of Physicists in Medicine, the European Association for Nuclear Medicine 11, the European Society of Radiology (ESR) 28, and the European Organisation for Research and Treatment of Cancer (EORTC) 55,86, among others. Where overlap does exists, the reference manual refers to existing recommendations and guidelines. This reference manual is divided into several chapters that describe processing of acquired and reconstructed (medical) imaging for high-throughput computation of image biomarkers (Chapter 2); that define a diverse set of image biomarkers (Chapter 3); that describe guidelines for report-

1 CHAPTER 1. INTRODUCTION 2 ing on radiomic studies and provide nomenclature for image biomarkers (Chapter 4); and that describe the data sets and image processing conﬁgurations used to ﬁnd reference values for image biomarkers (Chapter 5). Chapter 2

Image processing

Image processing is the sequence of operations required to derive image biomarkers (features) from acquired images. In the context of this work an image is defined as a three-dimensional (3D) stack of two-dimensional (2D) digital image slices. Image slices are stacked along the z-axis. This stack is furthermore assumed to possess the same coordinate system, i.e. image slices are not rotated or translated (in the xy-plane) with regard to each other. Moreover, digital images typically possess a finite resolution. Intensities in an image are thus located at regular intervals, or spacing. In 2D such regular positions are called pixels, whereas in 3D the term voxels is used. Pixels and voxels are thus represented as the intersections on a regularly spaced grid. Alternatively, pixels and voxels may be represented as rectangles and rectangular cuboids. The centers of the pixels and voxels then coincide with the intersections of the regularly spaced grid. Both representations are used in the document. Pixels and voxels contain an intensity value for each channel of the image. The number of channels depends on the imaging modality. Most medical imaging generates single-channel images, whereas the number of channels in microscopy may be greater, e.g. due to different stainings. In such multi-channel cases, features may be extracted for each separate channel, a subset of channels, or alternatively, channels may be combined and converted to a single-channel representation. In the remainder of the document we consider an image as if it only possesses a single channel. The intensity of a pixel or voxel is also called a grey level or grey tone, particularly in single- channel images. Though practically there is no difference, the terms grey level or grey tone are more commonly used to refer to discrete intensities, including discretised intensities. Image processing may be conducted using a wide variety of schemes. We therefore designed a general image processing scheme for image feature calculation based on schemes used within scientific literature 38. The image processing scheme is shown in figure 2.1. The processing steps referenced in the figure are described in detail within this chapter.

2.1 Data conversion 23XZ

Some imaging modalities require conversion of raw image data into a more meaningful presentation, e.g. standardised uptake values (SUV) 11. This is performed during the data conversion step. Assessment of data conversion methods falls outside the scope of the current work.

3 CHAPTER 2. IMAGE PROCESSING 4

Region of interest Image data ROI

Data conversion ROI interpolation

Image post-acquisition processing Intensity mask Morphological mask

Segmentation

Re-segmentation

VoxelImage interpolation interpolation

Calculation local intensity

Calculation morphological

ROI extraction Calculation statistical

Calculation Feature calculation IH, IVH*, GLCM, GLRLM GLSZM, NGTDM, NGLDM

Calculation Discretisation Feature data GLDZM

Figure 2.1 — Image processing scheme for image feature calculation. Depending on the specific imaging modality and purpose, some steps may be omitted. The region of interest (ROI) is explicitly split into two masks, namely an intensity and morphological mask, after interpolation to the same grid as the interpolated image. Feature calculation is expanded to show the different feature families with specific pre-processing. IH: intensity histogram; IVH: intensity-volume histogram; GLCM: grey level cooccurrence matrix; GLRLM: grey level run length matrix; GLSZM: grey level size zone matrix; NGTDM: neighbourhood grey tone difference matrix; NGLDM: Neighbouring grey level dependence matrix; GLDZM: grey level distance zone matrix; *Discretisation of IVH differs from IH and texture features, see section 3.5. CHAPTER 2. IMAGE PROCESSING 5

2.2 Image post-acquisition processing PCDE

Images are post-processed to enhance image quality. For instance, magnetic resonance imaging (MRI) contains both Gaussian and Rician noise 33 and may benefit from denoising. As another example, intensities measured using MR may be non-uniform across an image and could require correction 9,61,84. FDG-PET-based may furthermore be corrected for partial volume effects 12,66 and noise 26,43. In CT imaging, metal objects, e.g. pacemakers and tooth implants, introduce artifacts and may require artifinterpact suppression 32. Microscopy images generally benefit from field-of- view illumination correction as illumination is usually inhomogeneous due to the light-source or the optical path 13,62. Evaluation and standardisation of various image post-acquisition processing methods falls outside the scope of the current work. Note that vendors may provide or implement software to perform noise reduction and other post-processing during image reconstruction. In such cases, additional post-acquisition processing may not be required.

2.3 Segmentation OQYT

High-throughput image analysis, within the feature-based paradigm, relies on the definition of regions of interest (ROI). ROIs are used to define the region in which features are calculated. What constitutes an ROI depends on the imaging and the study objective. For example, in 3D microscopy of cell plates, cells are natural ROIs. In medical imaging of cancer patients, the tumour volume is a common ROI. ROIs can be defined manually by experts or (semi-)automatically using algorithms. From a process point-of-view, segmentation leads to the creation of an ROI mask R, for which every voxel j ∈ R (Rj) is defined as:

( 1 j in ROI Rj = 0 otherwise

ROIs are typically stored with the accompanying image. Some image formats directly store ROI masks as voxels (e.g. NIfTI, NRRD and DICOM Segmentation), and generating the ROI mask is conducted by loading the corresponding image. In other cases the ROI is saved as a set of (x, y, z) points that define closed loops of (planar) polygons, for example within DICOM RTSTRUCT or DICOM SR files. In such cases, we should determine which voxel centers lie within the space enclosed by the contour polygon in each slice to generate the ROI mask. A common method to determine whether a point in an image slice lies inside a 2D polygon is the crossing number algorithm, for which several implementations exist 58. The main concept behind this algorithm is that for any point inside the polygon, any line originating outside the polygon will cross the polygon an uneven number of times. A simple example is shown in figure 2.2. The implementation in the example makes use of the fact that the ROI mask is a regular grid to scan entire rows at a time. The example implementation consists of the following steps:

1. (optional) A ray is cast horizontally from outside the polygon for each of the n image rows. As we iterate over the rows, it is computationally beneﬁcial to exclude polygon edges that

will not be crossed by the ray for the current row j. If the current row has y-coordinate yj, and edge k has two vertices with y-coordinates yk1 and yk2, the ray will not cross the edge if both vertices lie either above or below yj, i.e. yj < yk1, yk2 or yj > yk1, yk2. For each row j, ﬁnd those polygon edges whose y-component of the vertices do not both lie on the same CHAPTER 2. IMAGE PROCESSING 6

side of the row coordinate yj. This step is used to limit calculation of intersection points to only those that cross a ray cast from outside the polygon – e.g. ray with origin (−1, yj) and direction (1, 0). This an optional step.

2. Determine intersection points xi of the (remaining) polygon edges with the ray.

3. Iterate over intersection points and add 1 to the count of each pixel center with x ≥ xi.

4. Apply the even-odd rule. Pixels with an odd count are inside the polygon, whereas pixels with an even count are outside.

Note that the example represents a relatively naive implementation that will not consistently assign voxel centers positioned on the polygon itself to the interior.

1. find intersecting polygons 3. count intersections along line grid with polygon contour

0 1 1 1 1 2

I II I II 2. find ray-polygon intersection 4. apply even-odd rule

I II I II

Figure 2.2 — Simple algorithm to determine which pixels are inside a 2D polygon. The suggested implementation consists of four steps: (1) Omit edges that will not intersect with the current row of voxel centers. (2) Calculate intersection points of edges I and II with the ray for the current row. (3) Determine the number of intersections crossed from ray origin to the row voxel centers. (4) Apply even-odd rule to determine whether voxel centers are inside the polygon.

2.4 Interpolation VTM2

Texture feature sets require interpolation to isotropic voxel spacing to be rotationally invariant, and to allow comparison between image data from different samples, cohorts or batches. Voxel interpolation affects image feature values as many image features are sensitive to changes in voxel size 4,8,59,60,87. Maintaining consistent isotropic voxel spacing across different measurements and devices is therefore important for reproducibility. At the moment there are no clear indications whether upsampling or downsampling schemes are preferable. Consider, for example, an image stack of slices with 1.0 × 1.0 × 3.0 mm3 voxel spacing. Upsampling to 1.0 × 1.0 × 1.0 mm3 requires inference and introduces artificial information, while conversely downsampling to the largest dimension (3.0×3.0×3.0 mm3) incurs information loss. Multiple-scaling strategies potentially offer a good trade-off 78. Note that downsampling may introduce image aliasing artifacts. Downsampling may therefore require anti-aliasing filters prior to filtering 49,90. While in general 3D interpolation algorithms are used to interpolate 3D images, 2D interpolation within the image slice plane may be recommended in some situations. In 2D interpolation voxels are not interpolated between slices. This may be beneficial if, for example, the spacing between slices is large compared to the desired voxel size, and/or compared to the in-plane spacing. Applying 3D interpolation would either require inferencing a large number of voxels between CHAPTER 2. IMAGE PROCESSING 7 slices (upsampling), or the loss of a large fraction of in-plane information (downsampling). The disadvantage of 2D interpolation is that voxel spacing is no longer isotropic, and as a consequence texture features can only be calculated in-plane.

Interpolation algorithms Interpolation algorithms translate image intensities from the original image grid to an interpolation grid. In such grids, voxels are spatially represented by their center. Several algorithms are commonly used for interpolation, such as nearest neighbour, trilinear, tricubic convolution and tricubic spline interpolation. In short, nearest neighbour interpolation assigns the intensity of the most nearby voxel in the original grid to each voxel in the interpolation grid. Trilinear interpolation uses the intensities of the eight most nearby voxels in the original grid to calculate a new interpolated intensity using linear interpolation. Tricubic convolution and tricubic spline interpolation draw upon a larger neighbourhood to evaluate a smooth, continuous third-order polynomial at the voxel centers in the interpolation grid. The difference between tricubic convolution and tricubic spline interpolation lies in the implementation. Whereas tricubic spline interpolation evaluates the smooth and continuous third-order polynomial at every voxel center, tricubic convolution approx- imates the solution using a convolution filter. Though tricubic convolution is faster, with modern hardware and common image sizes, the difference in execution speed is practically meaningless. Both interpolation algorithms produce similar results, and both are often referred to as tricubic interpolation. While no consensus exists concerning the optimal choice of interpolation algorithm, trilinear interpolation is usually seen as a conservative choice. It does not lead to the blockiness produced by nearest neighbour interpolation that introduces bias in local textures 38. Nor does it lead to out-of-range intensities which may occur due to overshoot with tricubic and higher order interpol- ations. The latter problem can occur in acute intensity transitions, where the local neighbourhood itself is not sufficiently smooth to evaluate the polynomial within the allowed range. Tricubic methods, however, may retain tissue contrast differences better. Particularly when upsampling, trilinear interpolation may act as a low-pass filter which suppresses higher spatial frequencies and cause artefacts in high-pass spatial filters. Interpolation algorithms and their advantages and disadvantages are treated in more detail elsewhere, e.g. Thévenaz et al. 70 . In a phantom study, Larue et al. 42 compared nearest neighbour, trilinear and tricubic interpolation and indicated that feature reproducibility is dependent on the selected interpolation algorithm, i.e. some features were more reproducible using one particular algorithm.

Rounding image intensities after interpolation 68QD Image intensities may require rounding after interpolation, or the application of cut-oﬀ values. For example, in CT images intensities represent Hounsﬁeld Units, and these do not take non-integer values. Following voxel interpolation, interpolated CT intensities are thus rounded to the nearest integer.

Partial volume effects in the ROI mask E8H9 If the image on which the ROI mask was defined, is interpolated after the ROI was segmented, the ROI mask R should likewise be interpolated to the same dimensions. Interpolation of the ROI mask is best conducted using either the nearest neighbour or trilinear interpolation methods, as these are guaranteed to produce meaningful masks. Trilinear interpolation of the ROI mask leads to partial volume effects, with some voxels containing fractions of the original voxels. Since a ROI CHAPTER 2. IMAGE PROCESSING 8 mask is a binary mask, such fractions need to be binarised by setting a partial volume threshold δ:

( 1 Rinterp,j ≥ δ Rj = 0 Rinterp,j < δ

A common choice for the partial volume threshold is δ = 0.5. For nearest neighbour interpolation the ROI mask does not contain partial volume fractions, and may be used directly. Interpolation results depend on the ﬂoating point representation used for the image and ROI masks. Floating point representations should at least be full precision (32-bit) to avoid rounding errors.

Interpolation grid UMPJ Interpolated voxel centers lie on the intersections of a regularly spaced grid. Grid intersections are represented by two coordinate systems. The first coordinate system is the grid coordinate system, with origin at (0.0, 0.0, 0.0) and distance between directly neighbouring voxel centers (spacing) of 1.0. The grid coordinate system is the coordinate system typically used by computers, and consequentially, by interpolation algorithms. The second coordinate system is the world coordinate system, which is typically found in (medical) imaging and provides an image scale. As the desired isotropic spacing is commonly defined in world coordinate dimensions, conversions between world coordinates and grid coordinates are necessary, and are treated in more detail after assessing grid alignment methods. Grid alignment affects feature values and is non-trivial. Three common grid alignments may be identified, and are shown in figure 2.3:

1. Fit to original grid (58MB). In this case the interpolation grid is deformed so that the voxel centers at the grid intersections overlap with the original grid vertices. For an original 4 × 4 voxel grid with spacing (3.00, 3.00) mm and a desired interpolation spacing of (2.00, 2.00) mm we first calculate the extent of the original voxel grid in world coordinates leading to an extent of ((4 − 1) 3.00, ((4 − 1) 3.00) = (9.00, 9.00) mm. In this case the interpolated grid will not exactly fit the original grid. Therefore we try to find the closest fitting grid, which leads to a 6 × 6 grid by rounding up (9.00/2.00, 9.00/2.00). The resulting grid has a grid spacing of (1.80, 1.80) mm in world coordinates, which differs from the desired grid spacing of (2.00, 2.00) mm.

2. Align grid origins (SBKJ). A simple approach which conserves the desired grid spacing is the alignment of the origins of the interpolation and original grids. Keeping with the same example, the interpolation grid is (6 × 6). The resulting voxel grid has a grid spacing of (2.00, 2.00) mm in world coordinates. By deﬁnition both grids are aligned at the origin, (0.00, 0.00).

3. Align grid centers (3WE3). The position of the origin may depend on image meta-data deﬁning image orientation. Not all software implementations may process this meta-data the same way. An implementation-independent solution is to align both grids on the grid center. Again, keeping with the same example, the interpolation grid is (6 × 6). Thus, the resulting voxel grid has a grid spacing of (2.00, 2.00) mm in world coordinates.

Align grid centers is recommended as it is implementation-independent and achieves the desired voxel spacing. Technical details of implementing align grid centers are described below.

Interpolation grid dimensions 026Q

The dimensions of the interpolation grid are determined as follows. Let na be the number of points along one axis of the original grid and sa,w their spacing in world coordinates. Then, let CHAPTER 2. IMAGE PROCESSING 9

sb,w be the desired spacing after interpolation. The axial dimension of the interpolated mesh grid is then: nasa nb = sb

Rounding towards inﬁnity guarantees that the interpolation grid exists even when the original grid contains few voxels. However, it also means that the interpolation mesh grid is partially located outside of the original grid. Extrapolation is thus required. Padding the original grid with the intensities at the boundary is recommended. Some implementations of interpolation algorithms may perform this padding internally.

Interpolation grid position QCY4 For the align grid centers method, the positions of the interpolation grid points are determined as follows. As before, let na and nb be the dimensions of one axis in the original and interpolation grid, respectively. Moreover, let sa,w be the original spacing and sb,w the desired spacing for the same axis in world coordinates. Then, with xa,w the origin of the original grid in world coordinates, the origin of the interpolation grid is located at:

s (n − 1) − s (n − 1) x = x + a a b b b,w a,w 2

In the grid coordinate system, the original grid origin is located at xa,g = 0. The origin of the interpolation grid is then located at: 1 sb,w xb,g = na − 1 − (nb − 1) 2 sa,w

Here the fraction sb,w/sa,w = sb,g is the desired spacing in grid coordinates. Thus, the interpolation grid points along the considered axis are located at grid coordinates:

xb,g, xb,g + sb,g, xb,g + 2sb,g, . . . , xb,g + (nb − 1)sb,g

Naturally, the above description applies to each grid axis. CHAPTER 2. IMAGE PROCESSING 10

original grid ﬁt to original grid

Original grid Interpola�on grid Size: 4x4 points Size: 6x6 points Spacing: (3.00, 3.00) Desired spacing: (2.00, 2.00) Realised spacing: (1.80, 1.80)

align grid origins align grid centers

Interpola�on grid Interpola�on grid Size: 6x6 points Size: 6x6 points Desired spacing: (2.00, 2.00) Desired spacing: (2.00, 2.00) Realised spacing: (2.00, 2.00) Realised spacing: (2.00, 2.00)

Figure 2.3 — Diﬀerent interpolation mesh grids based on an original 4 × 4 grid with (3.00, 3.00) mm spacing. The desired interpolation spacing is (2.00, 2.00) mm. Fit to original grid creates an interpolation mesh grid that overlaps with the corners of the original grid. Align grid origins creates an interpolation mesh grid that is positioned at the origin of the original grid. Align grid centers creates an interpolation grid that is centered on the center of original and interpolation grids. CHAPTER 2. IMAGE PROCESSING 11

2. ROI a�er re-segmenta�on 1. original ROI with range [1,6]

8 7 6 3 8 7 6 3 1 1 1 2 1 2 1 1 1 2 1 2 1 2 2 2 2 1 1 1 2 2 2 2 1 1 1 4 8 8 1 1 4 8 8 1 3 7 5 6 7 8 3 7 5 6 7 8 1 1 4 3 7 7 8 1 1 4 3 7 7 8 2 8 7 4 1 8 1 2 8 7 4 1 8 1 1 1 1 3 1 1 1 1 3 1

3a. re-segmented ROI 3b. re-segmented ROI (morphological mask) (intensity mask)

6 3 1 1 1 2 1 2 1 2 2 2 2 1 1 1 4 1 3 5 6 1 1 4 3 2 4 1 1 1 1 1 3 1

outside ROI inside ROI re-segmented pixel

Figure 2.4 — Example showing how intensity and morphological masks may diﬀer due to re- segmentation. (1) The original region of interest (ROI) is shown with pixel intensities. (2) Sub- sequently, the ROI is re-segmented to only contain values in the range [1, 6]. Pixels outside this range are marked for removal from the intensity mask. (3a) Resulting morphological mask, which is identical to the original ROI. (3b) Re-segmented intensity mask. Note that due to re-segmentation, intensity and morphological masks are diﬀerent.

2.5 Re-segmentation IF9H

Re-segmentation entails updating the ROI mask R based on corresponding voxel intensities Xgl. Re-segmentation may be performed to exclude voxels from a previously segmented ROI, and is performed after interpolation. An example use would be the exclusion of air or bone voxels from an ROI deﬁned on CT imaging. Two common re-segmentation methods are described in this section. Combining multiple re-segmentation methods is possible. In this case, the intersection of the intensity ranges deﬁned by the re-segmentation methods is used.

Intensity and morphological masks of an ROI ECJF Conventionally, an ROI consists of a single mask. However, re-segmentation may lead to exclusion of internal voxels, or divide the ROI into sub-volumes. To avoid undue complexity by again updating the re-segmented ROI for a more plausible morphology, we deﬁne two separate ROI masks. The morphological mask (G5KJ) is not re-segmented and maintains the original morphology as deﬁned by an expert and/or (semi-)automatic segmentation algorithms. CHAPTER 2. IMAGE PROCESSING 12

CT image slice with mask range: [-50, 50] HU

outlier: 1.0 σ range: [-50, 50] HU outlier: 1.0 σ Figure 2.5 — Re-segmentation example based on a CT-image. The masked region (blue) is re- segmented to create an intensity mask (orange). Examples using three diﬀerent re-segmentation parameter sets are shown. The bottom right combines the range and outlier re-segmentation, and the resulting mask is the intersection of the masks in the other two examples. Image data from Valli`eres et al. 16,76,77.

The intensity mask (SEFI) can be re-segmented and will contain only the selected voxels. For many feature families, only this is important. However, for morphological and grey level distance zone matrix (GLDZM) feature families, both intensity and morphological masks are used. A two-dimensional schematic example is shown in ﬁgure 2.4, and a real example is shown in ﬁgure 2.5.

Range re-segmentation USB3 Re-segmentation may be performed to remove voxels from the intensity mask that fall outside of a specified range. An example is the exclusion of voxels with Hounsfield Units indicating air and bone tissue in the tumour ROI within CT images, or low activity areas in PET images. Such ranges of intensities of included voxels are usually presented as a closed interval [a, b] or half-open interval [a, ∞), respectively. For arbitrary intensity units (found in e.g. raw MRI data, uncalibrated microscopy images, and many spatial filters), no re-segmentation range can be provided. When a re-segmentation range is defined by the user, it needs to be propagated and used for the calculation of features that require a specified intensity range (e.g. intensity-volume histogram features) and/or that employs fixed bin size discretisation. Recommendations for the possible combinations of different imaging intensity definitions, re-segmentation ranges and discretisation algorithms are provided in Table 2.1. CHAPTER 2. IMAGE PROCESSING 13

Intensity outlier filtering 7ACA ROI voxels with outlier intensities may be removed from the intensity mask. One method for defining outliers was suggested by Vallièreset al. 76 after Collewet et al. 19 . The mean µ and standard deviation σ of grey levels of voxels assigned to the ROI are calculated. Voxels outside the range [µ − 3σ, µ + 3σ] are subsequently excluded from the intensity mask.

2.6 ROI extraction 1OBP

Many feature families require that the ROI is isolated from the surrounding voxels. The ROI intensity mask is used to extract the image volume to be studied. Excluded voxels are commonly replaced by a placeholder value, often NaN. This placeholder value may then used to exclude these voxels from calculations. Voxels included in the ROI mask retain their original intensity. An example is shown in ﬁgure 2.6.

2.7 Intensity discretisation 4R0B

Discretisation or quantisation of image intensities inside the ROI is often required to make calculation of texture features tractable 88, and possesses noise-suppressing properties as well. An example of discretisation is shown in figure 2.7. Two approaches to discretisation are commonly used. One involves the discretisation to a fixed number of bins, and the other discretisation with a fixed bin width. As we will observe, there is no inherent preference for one or the other method. However, both methods have particular characteristics (as described below) that may make them better suited for specific purposes. Note that the lowest bin always has value 1, and not 0. This ensures consistency for calculations of texture features, where for some features grey level 0 is not allowed .

Fixed bin number K15C

In the fixed bin number method, intensities Xgl are discretised to a fixed number of Ng bins. It is defined as follows:  j Xgl,k−Xgl,min k  Ng + 1 Xgl,k < Xgl,max Xgl,max−Xgl,min Xd,k = Ng Xgl,k = Xgl,max

In short, the intensity Xgl,k of voxel k is corrected by the lowest occurring intensity Xgl,min in the ROI, divided by the bin width (Xgl,max − Xgl,min) /Ng, and subsequently rounded down to the

apply mask

Figure 2.6 — Masking of an image by the ROI mask during ROI extraction. Intensities outside the ROI are excluded. Image data from Valli`ereset al. 16,76,77. CHAPTER 2. IMAGE PROCESSING 14

ﬁxed bin number discretisation

image A

image B

original 32 bins 16 bins 8 bins 4 bins

ﬁxed bin size discretisation

image A

image B

original 0.25 SUV 0.50 SUV 1.00 SUV 2.00 SUV

Figure 2.7 — Discretisation of two different 18F-FDG-PET images with SUVmax of 27.8 (A) and 6.6 (B). Fixed bin number discretisation adjust the contrast between the two images, with the number of bins determining the coarseness of the discretised image. Fixed bin size discretisation leaves the contrast differences between image A and B intact. Increasing the bin size increases the coarseness of the discretised image. Image data from Vallièreset al. 16,76,77. nearest integer (floor function). The fixed bin number method breaks the relationship between image intensity and physiological meaning (if any). However, it introduces a normalising effect which may be beneficial when intensity units are arbitrary (e.g. raw MRI data and many spatial filters), and where contrast is considered important. Furthermore, as values of many features depend on the number of grey levels found within a given ROI, the use of a fixed bin number discretisation algorithm allows for a direct comparison of feature values across multiple analysed ROIs (e.g. across different samples).

Fixed bin size Q3RU Fixed bin size discretisation is conceptually simple. A new bin is assigned for every intensity interval with width wb; i.e. wb is the bin width, starting at a minimum Xgl,min. The minimum intensity may be a user-set value as defined by the lower bound of the re-segmentation range, or data-driven as defined by the minimum intensity in the ROI Xgl,min = min (Xgl). In all cases, the method used and/or set minimum value must be clearly reported. However, to maintain consistency between samples, we strongly recommend to always set the same minimum value for all samples as defined by the lower bound of the re-segmentation range (e.g. HU of -500 for CT, SUV of 0 for PET, etc.). In the case that no re-segmentation range may be defined due to arbitrary intensity units (e.g. raw MRI data and many spatial filters), the use of the fixed bin size discretisation algorithm is not recommended. The fixed bin size method has the advantage of maintaining a direct relationship with the original intensity scale, which could be useful for functional imaging modalities such as PET. CHAPTER 2. IMAGE PROCESSING 15

Imaging intensity Re-segmentation FBN(2) FBS(3) units(1) range [a, b] calibrated [a, ∞) none

arbitrary none

Table 2.1 — Recommendations for the possible combinations of different imaging intensity definitions, re-segmentation ranges and discretisation algorithms. Checkmarks () represent recommended combinations of re-segmentation range and discretisation algorithm, whereas crossmarks () represent non-recommended combinations. (1) PET and CT are examples of imaging modalities with calibrated intensity units (e.g. SUV and HU, respectively), and raw MRI data of arbitrary intensity units. (2) Fixed bin number (FBN) discretisation uses the actual range of intensities in the analysed ROI (re-segmented or not), and not the re-segmentation range itself (when defined). (3) Fixed bin size (FBS) discretisation uses the lower bound of the re-segmentation range as the minimum set value. When the re-segmentation range is not or cannot be defined (e.g. arbitrary intensity units), the use of the FBS algorithm is not recommended.

Discretised intensities are computed as follows: Xgl,k − Xgl,min Xd,k = + 1 wb

In short, the minimum intensity Xgl,min is subtracted from intensity Xgl,k in voxel k, and then divided by the bin width wb. The resulting value is subsequently rounded down to the nearest integer (ﬂoor function), and 1 is added to arrive at the discretised intensity.

Other methods Many other methods and variations for discretisation exist, but are not described in detail here. Vallièreset al. 76 described the use of intensity histogram equalisation and Lloyd-Max algorithms for discretisation. Intensity histogram equalisation involves redistributing intensities so that the resulting bins contain a similar number of voxels, i.e. contrast is increased by flattening the histogram as much as possible 34. Histogram equalisation of the ROI imaging intensities can be performed before any other discretisation algorithm (e.g. FBN, FSB, etc.), and it also requires the definition of a given number of bins in the histogram to be equalised. The Lloyd-Max algorithm is an iterative clustering method that seeks to minimise mean squared discretisation errors 47,50.

Recommendations The discretisation method that leads to optimal feature inter- and intra-sample reproducibility is modality-dependent. Usage recommendations for the possible combinations of diﬀerent imaging intensity deﬁnitions, re-segmentation ranges and discretisation algorithms are provided in Table 2.1. Overall, the discretisation choice has a substantial impact on intensity distributions, feature values and reproducibility 4,25,37,38,44,59,83.

2.8 Feature calculation

Feature calculation is the final processing step where feature descriptors are used to quantify characteristics of the ROI. After calculation such features may be used as image biomarkers by CHAPTER 2. IMAGE PROCESSING 16 relating them to physiological and medical outcomes of interest. Feature calculation is handled in full details in the next chapter. Let us recall that the image processing steps leading to image biomarker calculations can be performed in many different ways, notably in terms of spatial filtering, segmentation, interpolation and discretisation parameters. Furthermore, it is plausible that different texture features will better quantify the characteristics of the ROI when computed using different image processing parameters. For example, a lower number of grey levels in the discretisation process (e.g. 8 or 16) may allow to better characterize the sub-regions of the ROI using grey level size zone matrix (GLSZM) features, whereas grey level co-occurence matrix (GLCM) features may be better modeled with a higher number of grey levels (e.g. 32 or 64). Overall, these possible differences opens the door to the optimization of image processing parameters for each different feature in terms of a specific objective. For the specific case of the optimization of image interpolation and discretisation prior to texture analysis, Vallières et al. 76 have named this process texture optimization. The authors notably suggested that the texture optimization process could have significant influence of the prognostic capability of subsequent features. In another study 78, the authors constructed predictive models using textures calculated from all possible combinations of PET and CT images interpolated at four isotropic resolutions and discretised with two different algorithms and four numbers of grey levels. Chapter 3

Image features

In this chapter we will describe a set of quantitative image features together with the reference values established by the IBSI. This feature set builds upon the feature sets proposed by Aerts et al. 1 and Hatt et al. 38 , which are themselves largely derived from earlier works. References to earlier work are provided whenever they could be identified. Reference values were derived for each feature. A table of reference values contains the values that could be reliably reproduced, within a tolerance margin, for the reference data sets (see Chapter 5). Consensus on the validity of each reference value is also noted. Consensus can have four levels, depending on the number of teams that were able to produce the same value during the standardization process: weak (< 3 matches), moderate (3 to 5 matches), strong (6 to 9 matches), and very strong (≥ 10 matches). If consensus on a reference value was weak or if it could not be reproduced by an absolute majority of teams, it was not considered standardized. Such features do currently not have reference values, and should not be used. The set of features can be divided into a number of families, of which intensity-based statistical, intensity histogram-based, intensity-volume histogram-based, morphological features, local intensity, and texture matrix-based features are treated here. All texture matrices are rotationally and translationally invariant. Illumination invariance of texture matrices may be achieved by particular image post-acquisition schemes, e.g. histogram matching. None of the texture matrices are scale invariant, a property which can be useful in many (biomedical) applications. What the presented texture matrices lack, however, is directionality in combination with rotation invariance. These may be achieved by local binary patterns and steerable filters, which however fall beyond the scope of the current work. For these and other texture features, see Depeursinge et al. 24 . Features are calculated on the base image, as well as images transformed using wavelet or Gabor filters). To calculate features, it is assumed that an image segmentation mask exists, which identifies the voxels located within a region of interest (ROI). The ROI itself consists of two masks, an intensity mask and a morphological mask. These masks may be identical, but not necessarily so, as described in Section 2.5. Several feature families require additional image processing steps before feature calculation. Notably intensity histogram and texture feature families require prior discretisation of intensities into grey level bins. Other feature families do not require discretisation before calculations. For more details on image processing, see figure 2.1 in the previous chapter. Below is an overview table that summarises image processing requirements for the different feature families.

17 CHAPTER 3. IMAGE FEATURES 18

ROI mask Feature family count morph. int. discr. morphology 29 local intensity 2 a intensity-based statistics 18 intensity histogram 23 intensity-volume histogram 5 b grey level co-occurrence matrix 25 grey level run length matrix 16 grey level size zone matrix 16 grey level distance zone matrix 16 neighbourhood grey tone diﬀerence matrix 5 neighbouring grey level dependence matrix 17

Table 3.1 — Feature families and required image processing. For each feature family, the number of features in the document, the required input of a morphological (morph.) and/or intensity (int.) ROI mask, as well as the requirement of image discretisation (discr.) is provided. a The entire image volume should be available when computing local intensity features. b Image discretisation for the intensity-volume histogram is performed with ﬁner discretisation than required for e.g. textural features.

Though image processing parameters aﬀect feature values, three other concepts inﬂuence feature values for many features: distance, feature aggregation and distance weighting. These are described below.

Grid distances MPUJ Grid distance is an important concept that is used by several feature families, particularly texture features. Grid distances can be measured in several ways. Let m = (mx, my, mz) be the vector from a center voxel at k = (kx, ky, kz) to a neighbour voxel at k + m. The following norms (distances) are used:

• `1 norm or Manhattan norm (LIFZ):

kmk1 = |mx| + |my| + |mz|

• `2 norm or Euclidean norm (G9EV): q 2 2 2 kmk2 = mx + my + mz

• `∞ norm or Chebyshev norm (PVMT):

kmk∞ = max(|mx|, |my|, |mz|)

An example of how the above norms diﬀer in practice is shown in ﬁgure 3.1. CHAPTER 3. IMAGE FEATURES 19

(a) Manhattan norm (b) Euclidean norm (c) Chebyshev norm

Figure 3.1 — Grid neighbourhoods for distances up to 3 according to Manhattan, Euclidean and Chebyshev norms. The orange pixel is considered the center pixel. Dark blue pixels have distance δ = 1, blue pixels δ ≤ 2 and light blue pixels δ ≤ 3 for the corresponding norm.

Feature aggregation 5QB6 Features from some families may be calculated from, e.g. slices. As a consequence, multip le values for the same feature may be computed. These diﬀerent values should be combined into a single value for many common purposes. This process is referred to as feature aggregation. Feature aggregation methods depend on the family, and are detailed in the family description.

Distance weighting 6CK8 Distance weighting is not a default operation for any of the texture families, but is implemented in software such as PyRadiomics 81. It may for example be used to put more emphasis on local intensities. CHAPTER 3. IMAGE FEATURES 20

3.1 Morphological features HCUG

Morphological features describe geometric aspects of a region of interest (ROI), such as area and volume. Morphological features are based on ROI voxel representations of the volume. Three voxel representations of the volume are conceivable:

1. The volume is represented by a collection of voxels with each voxel taking up a certain volume (LQD8).

2. The volume is represented by a voxel point set Xc that consists of coordinates of the voxel centers (4KW8).

3. The volume is represented by a surface mesh (WRJH).

We use the second representation when the inner structure of the volume is important, and the third representation when only the outer surface structure is important. The ﬁrst representation is not used outside volume approximations because it does not handle partial volume eﬀects at the ROI edge well, and also to avoid inconsistencies in feature values introduced by mixing representations in small voxel volumes.

Mesh-based representation WRJH A mesh-based representation of the outer surface allows consistent evaluation of the surface volume and area independent of size. Voxel-based representations lead to partial volume effects and over-estimation of the surface area. The surface of the ROI volume is translated into a triangle mesh using a meshing algorithm. While multiple meshing algorithms exist, we suggest the use of the Marching Cubes algorithm 45,48 because of its widespread availability in different programming languages and reasonable approximation of the surface area and volume 67. In practice, mesh- based feature values depend upon the meshing algorithm and small differences may occur between implementations 46.

b a

Figure 3.2 — Meshing algorithms draw faces and vertices to cover the ROI. One face, spanned by vertices a, b and c, is highlighted. Moreover, the vertices deﬁne the three edges ab = b − a, bc = c − b and ca = a − c. The face normal n is determined using the right-hand rule, and calculated as n = (ab × bc) /kab × bck, i.e. the outer product of edge ab with edge bc, normalised by its length.

Meshing algorithms use the ROI voxel point set Xc to create a closed mesh. Dependent on the algorithm, a parameter is required to specify where the mesh should be drawn. A default level of 0.5 times the voxel spacing is used for marching cube algorithms. Other algorithms require a so-called isovalue, for which a value of 0.5 can be used since the ROI mask consists of 0 and 1 values, and we want to roughly draw the mesh half-way between voxel centers. Depending on implementation, algorithms may also require padding of the ROI mask with non-ROI (0) voxels CHAPTER 3. IMAGE FEATURES 21 to correctly estimate the mesh in places where ROI voxels would otherwise be located at the edge of the mask.

The closed mesh drawn by the meshing algorithm consists of Nfc triangle faces spanned by Nvx vertex points. An example triangle face is drawn in Figure 3.2. The set of vertex points is then Xvx. The calculation of the mesh volume requires that all faces have the same orientation of the face normal. Consistent orientation can be checked by the fact that in a regular, closed mesh, all edges are shared between exactly two faces. Given the edge spanned by vertices a and b, the edge must be ab = b − a for one face and ba = a − b for the adjacent face. This ensures consistent application of the right-hand rule, and thus consistent orientation of the face normals. Algorithm implementations may return consistently orientated faces by default.

ROI morphological and intensity masks The ROI consists of a morphological and an intensity mask. The morphological mask is used to calculate many of the morphological features and to generate the voxel point set Xc. Any holes within the morphological mask are understood to be the result of segmentation decisions, and thus to be intentional. The intensity mask is used to generate the voxel intensity set Xgl with corresponding point set Xc,gl.

Aggregating features By deﬁnition, morphological features are calculated in 3D (DHQ4), and not per slice.

Units of measurement By definition, morphological features are computed using the unit of length as defined in the DICOM standard, i.e. millimeter for most medical imaging modalities1. If the unit of length is not defined by a standard, but is explicitly defined as meta data, this definition should be used. In this case, care should be taken that this definition is consistent across all data in the cohort. If a feature value should be expressed as a different unit of length, e.g. cm instead of mm, such conversions should take place after computing the value using the standard units.

3.1.1 Volume (mesh) RNU0

The mesh-based volume V is calculated from the ROI mesh as follows 89. A tetrahedron is formed by each face k and the origin. By placing the origin vertex of each tetrahedron at (0, 0, 0), the signed volume of the tetrahedron is:

a · (b × c) V = k 6 Here a, b and c are the vertex points of face k. Depending on the orientation of the normal, the signed volume may be positive or negative. Hence, the orientation of face normals should be consistent, e.g. all normals must be either pointing outward or inward. The volume V is then calculated by summing over the face volumes, and taking the absolute value:

Nfc X Fmorph.vol = V = Vk

k=1

1DICOM PS3.3 2019a - Information Object Deﬁnitions, Section 10.7.1.3 CHAPTER 3. IMAGE FEATURES 22

In positron emission tomography, the volume of the ROI commonly receives a name related to the radioactive tracer, e.g. metabolically active tumour volume (MATV) for 18F-FDG.

data value tol. consensus

dig. phantom 556 4 very strong config. A 3.58 × 105 5 × 103 very strong config. B 3.58 × 105 5 × 103 strong config. C 3.67 × 105 6 × 103 strong config. D 3.67 × 105 6 × 103 strong config. E 3.67 × 105 6 × 103 strong

Table 3.2 — Reference values for the volume (mesh) feature.

3.1.2 Volume (voxel counting) YEKZ

In clinical practice, volumes are commonly determined by counting voxels. For volumes consisting of a large number of voxels (1000s), the diﬀerences between voxel counting and mesh-based approaches are usually negligible. However for volumes with a low number of voxels (10s to 100s), voxel counting will overestimate volume compared to the mesh-based approach. It is therefore only used as a reference feature, and not in the calculation of other morphological features. Voxel counting volume is deﬁned as:

N Xv Fmorph.approx.vol = Vk k=1

Here Nv is the number of voxels in the morphological mask of the ROI, and Vk the volume of voxel k.

data value tol. consensus

dig. phantom 592 4 very strong config. A 3.59 × 105 5 × 103 strong config. B 3.58 × 105 5 × 103 strong config. C 3.68 × 105 6 × 103 strong config. D 3.68 × 105 6 × 103 strong config. E 3.68 × 105 6 × 103 strong

Table 3.3 — Reference values for the volume (voxel counting) feature.

3.1.3 Surface area (mesh) C0JK

The surface area A is also calculated from the ROI mesh by summing over the triangular face surface areas 1. By deﬁnition, the area of face k is:

|ab × ac| A = k 2 CHAPTER 3. IMAGE FEATURES 23

As in Figure 3.2, edge ab = b − a is the vector from vertex a to vertex b, and edge ac = c − a the vector from vertex a to vertex c. The total surface area A is then:

Nfc X Fmorph.area = A = Ak k=1

data value tol. consensus

dig. phantom 388 3 very strong config. A 3.57 × 104 300 strong config. B 3.37 × 104 300 strong config. C 3.43 × 104 400 strong config. D 3.43 × 104 400 strong config. E 3.43 × 104 400 strong

Table 3.4 — Reference values for the surface area (mesh) feature.

3.1.4 Surface to volume ratio 2PR5

The surface to volume ratio is given as 1:

A F = morph.av V

Note that this feature is not dimensionless.

data value tol. consensus

dig. phantom 0.698 0.004 very strong config. A 0.0996 0.0005 strong config. B 0.0944 0.0005 strong config. C 0.0934 0.0007 strong config. D 0.0934 0.0007 strong config. E 0.0934 0.0007 strong

Table 3.5 — Reference values for the surface to volume ratio feature.

3.1.5 Compactness 1 SKGS

Several features (compactness 1 and 2, spherical disproportion, sphericity and asphericity) quantify the deviation of the ROI volume from a representative spheroid. All these deﬁnitions can be derived from one another. As a results these features are are highly correlated and may thus be redundant. Compactness 1 1 is a measure for how compact, or sphere-like the volume is. It is deﬁned as:

V F = morph.comp.1 π1/2A3/2

Compactness 1 is sometimes 1 deﬁned using A2/3 instead of A3/2, but this does not lead to a dimensionless quantity. CHAPTER 3. IMAGE FEATURES 24

data value tol. consensus

dig. phantom 0.0411 0.0003 strong config. A 0.03 0.0001 strong config. B 0.0326 0.0001 strong config. C — — moderate config. D 0.0326 0.0002 strong config. E 0.0326 0.0002 strong

Table 3.6 — Reference values for the compactness 1 feature. An unset value (—) indicates the lack of a reference value.

3.1.6 Compactness 2 BQWJ

Like Compactness 1, Compactness 2 1 quantiﬁes how sphere-like the volume is:

V 2 F = 36π morph.comp.2 A3

1/2 By deﬁnition Fmorph.comp.1 = 1/6π (Fmorph.comp.2 ) .

data value tol. consensus

dig. phantom 0.599 0.004 strong config. A 0.319 0.001 strong config. B 0.377 0.001 strong config. C 0.378 0.004 strong config. D 0.378 0.004 strong config. E 0.378 0.004 strong

Table 3.7 — Reference values for the compactness 2 feature.

3.1.7 Spherical disproportion KRCK

Spherical disproportion 1 likewise describes how sphere-like the volume is:

A A Fmorph.sph.dispr = = 4πR2 (36πV 2)1/3

−1/3 By deﬁnition Fmorph.sph.dispr = (Fmorph.comp.2 ) .

data value tol. consensus

dig. phantom 1.19 0.01 strong config. A 1.46 0.01 strong config. B 1.38 0.01 strong config. C 1.38 0.01 strong config. D 1.38 0.01 strong config. E 1.38 0.01 strong

Table 3.8 — Reference values for the spherical disproportion feature. CHAPTER 3. IMAGE FEATURES 25

3.1.8 Sphericity QCFX

Sphericity 1 is a further measure to describe how sphere-like the volume is:

1/3 36πV 2 F = morph.sphericity A

1/3 By deﬁnition Fmorph.sphericity = (Fmorph.comp.2 ) .

data value tol. consensus

dig. phantom 0.843 0.005 very strong config. A 0.683 0.001 strong config. B 0.722 0.001 strong config. C 0.723 0.003 strong config. D 0.723 0.003 strong config. E 0.723 0.003 strong

Table 3.9 — Reference values for the sphericity feature.

3.1.9 Asphericity 25C7

Asphericity 6 also describes how much the ROI deviates from a perfect sphere, with perfectly spherical volumes having an asphericity of 0. Asphericity is deﬁned as:

1 A3 1/3 F = − 1 morph.asphericity 36π V 2

−1/3 By deﬁnition Fmorph.asphericity = (Fmorph.comp.2 ) − 1

data value tol. consensus

dig. phantom 0.186 0.001 strong config. A 0.463 0.002 strong config. B 0.385 0.001 moderate config. C 0.383 0.004 strong config. D 0.383 0.004 strong config. E 0.383 0.004 strong

Table 3.10 — Reference values for the asphericity feature.

3.1.10 Centre of mass shift KLMA

The distance between the ROI volume centroid and the intensity-weighted ROI volume is an abstraction of the spatial distribution of low/high intensity regions within the ROI. Let Nv,m be the number of voxels in the morphological mask. The ROI volume centre of mass is calculated from the morphological voxel point set Xc as follows:

Nv,m −−−→ 1 X CoM geom = X~ c,k Nv,m k=1 CHAPTER 3. IMAGE FEATURES 26

The intensity-weighted ROI volume is based on the intensity mask. The position of each voxel centre in the intensity mask voxel set Xc,gl is weighted by its corresponding intensity Xgl. There- fore, with Nv,gl the number of voxels in the intensity mask:

PNv,gl ~ −−−→ k=1 Xgl,kXc,gl,k CoM gl = PNv,gl k=1 Xgl,k

The distance between the two centres of mass is then: −−−→ −−−→ Fmorph.com = ||CoM geom − CoM gl||2

data value tol. consensus

dig. phantom 0.672 0.004 very strong config. A 52.9 28.7 strong config. B 63.1 29.6 strong config. C 45.6 2.8 strong config. D 64.9 2.8 strong config. E 68.5 2.1 moderate

Table 3.11 — Reference values for the centre of mass shift feature.

3.1.11 Maximum 3D diameter L0JK

The maximum 3D diameter 1 is the distance between the two most distant vertices in the ROI mesh vertex set Xvx:

~ ~ Fmorph.diam = max ||Xvx,k1 − Xvx,k2 ||2 , k1 = 1, . . . , N k2 = 1,...,N

A practical way of determining the maximum 3D diameter is to first construct the convex hull of the ROI mesh. The convex hull vertex set Xvx,convex is guaranteed to contain the two most distant vertices of Xvx. This significantly reduces the computational cost of calculating distances between all vertices. Despite the remaining O(n2) cost of calculating distances between different vertices, Xvx,convex is usually considerably smaller in size than Xvx. Moreover, the convex hull is later used for the calculation of other morphological features (3.1.25-3.1.26).

data value tol. consensus

dig. phantom 13.1 0.1 strong config. A 125 1 strong config. B 125 1 strong config. C 125 1 strong config. D 125 1 strong config. E 125 1 strong

Table 3.12 — Reference values for the maximum 3D diameter feature.

3.1.12 Major axis length TDIC

Principal component analysis (PCA) can be used to determine the main orientation of the ROI 65.

On a three dimensional object, PCA yields three orthogonal eigenvectors {e1, e2, e3} and three CHAPTER 3. IMAGE FEATURES 27

eigenvalues (λ1, λ2, λ3). These eigenvalues and eigenvectors geometrically describe a triaxial ellipsoid. The three eigenvectors determine the orientation of the ellipsoid, whereas the eigenvalues provide a measure of how far the ellipsoid extends along each eigenvector. Several features make use of principal component analysis, namely major, minor and least axis length, elongation, ﬂatness, and approximate enclosing ellipsoid volume and area density.

The eigenvalues can be ordered so that λmajor ≥ λminor ≥ λleast correspond to the major, minor and least axes of the ellipsoid respectively. The semi-axes lengths a, b and c for the major, p √ √ minor and least axes are then 2 λmajor , 2 λminor and 2 λleast respectively. The major axis length is twice the semi-axis length a, determined using the largest eigenvalue obtained by PCA 39 on the point set of voxel centers Xc : p Fmorph.pca.major = 2a = 4 λmajor

data value tol. consensus

dig. phantom 11.4 0.1 very strong config. A 92.7 0.4 very strong config. B 92.6 0.4 strong config. C 93.3 0.5 strong config. D 93.3 0.5 strong config. E 93.3 0.5 strong

Table 3.13 — Reference values for the major axis length feature.

3.1.13 Minor axis length P9VJ

The minor axis length of the ROI provides a measure of how far the volume extends along the second largest axis. The minor axis length is twice the semi-axis length b, determined using the second largest eigenvalue obtained by PCA, as described in Section 3.1.12:

p Fmorph.pca.minor = 2b = 4 λminor

data value tol. consensus

dig. phantom 9.31 0.06 very strong config. A 81.5 0.4 very strong config. B 81.3 0.4 strong config. C 82 0.5 strong config. D 82 0.5 strong config. E 82 0.5 strong

Table 3.14 — Reference values for the minor axis length feature.

3.1.14 Least axis length 7J51

The least axis is the axis along which the object is least extended. The least axis length is twice the semi-axis length c, determined using the smallest eigenvalue obtained by PCA, as described in CHAPTER 3. IMAGE FEATURES 28

Section 3.1.12:

p Fmorph.pca.least = 2c = 4 λleast

data value tol. consensus

dig. phantom 8.54 0.05 very strong config. A 70.1 0.3 strong config. B 70.2 0.3 strong config. C 70.9 0.4 strong config. D 70.9 0.4 strong config. E 70.9 0.4 strong

Table 3.15 — Reference values for the least axis length feature.

3.1.15 Elongation Q3CK

The ratio of the major and minor principal axis lengths could be viewed as the extent to which a volume is longer than it is wide, i.e. is eccentric. For computational reasons, we express elongation as an inverse ratio. 1 is thus completely non-elongated, e.g. a sphere, and smaller values express greater elongation of the ROI volume. s λminor Fmorph.pca.elongation = λmajor

data value tol. consensus

dig. phantom 0.816 0.005 very strong config. A 0.879 0.001 strong config. B 0.878 0.001 strong config. C 0.879 0.001 strong config. D 0.879 0.001 strong config. E 0.879 0.001 strong

Table 3.16 — Reference values for the elongation feature.

3.1.16 Flatness N17B

The ratio of the major and least axis lengths could be viewed as the extent to which a volume is flat relative to its length. For computational reasons, we express flatness as an inverse ratio. 1 is thus completely non-flat, e.g. a sphere, and smaller values express objects which are increasingly flatter. s λleast Fmorph.pca.flatness = λmajor CHAPTER 3. IMAGE FEATURES 29

data value tol. consensus

dig. phantom 0.749 0.005 very strong config. A 0.756 0.001 strong config. B 0.758 0.001 strong config. C 0.76 0.001 strong config. D 0.76 0.001 strong config. E 0.76 0.001 strong

Table 3.17 — Reference values for the ﬂatness feature.

3.1.17 Volume density (axis-aligned bounding box) PBX1

Volume density is the fraction of the ROI volume and a comparison volume. Here the comparison volume is that of the axis-aligned bounding box (AABB) of the ROI mesh vertex set Xvx or the ROI mesh convex hull vertex set Xvx,convex. Both vertex sets generate an identical bounding box, which is the smallest box enclosing the vertex set, and aligned with the axes of the reference frame.

V Fmorph.v.dens.aabb = Vaabb

This feature is also called extent 27,65.

data value tol. consensus

dig. phantom 0.869 0.005 strong config. A 0.486 0.003 strong config. B 0.477 0.003 strong config. C 0.478 0.003 strong config. D 0.478 0.003 strong config. E 0.478 0.003 strong

Table 3.18 — Reference values for the volume density (AABB) feature.

3.1.18 Area density (axis-aligned bounding box) R59B

Conceptually similar to the volume density (AABB) feature, area density considers the ratio of the

ROI surface area and the surface area Aaabb of the axis-aligned bounding box enclosing the ROI 80 mesh vertex set Xvx . The bounding box is identical to the one used for computing the volume density (AABB) feature. Thus:

A Fmorph.a.dens.aabb = Aaabb CHAPTER 3. IMAGE FEATURES 30

data value tol. consensus

dig. phantom 0.866 0.005 strong config. A 0.725 0.003 strong config. B 0.678 0.003 strong config. C 0.678 0.003 strong config. D 0.678 0.003 strong config. E 0.678 0.003 strong

Table 3.19 — Reference values for the area density (AABB) feature.

3.1.19 Volume density (oriented minimum bounding box) ZH1A

Note: This feature currently has no reference values and should not be used. The volume of an axis-aligned bounding box is generally not the smallest obtainable volume enclosing the ROI. By orienting the box along a diﬀerent set of axes, a smaller enclosing volume may be attainable. The oriented minimum bounding box (OMBB) of the ROI mesh vertex set

Xvx or Xvx,convex encloses the vertex set and has the smallest possible volume. A 3D rotating callipers technique was devised by O’Rourke 56 to derive the oriented minimum bounding box. Due to computational complexity of this technique, the oriented minimum bounding box is commonly approximated at lower complexity, see e.g. Barequet and Har-Peled 10 and Chan and Tan 14 . Thus:

V Fmorph.v.dens.ombb = Vombb

Here Vombb is the volume of the oriented minimum bounding box.

3.1.20 Area density (oriented minimum bounding box) IQYR

Note: This feature currently has no reference values and should not be used. The area density (OMBB) is estimated as:

A Fmorph.a.dens.ombb = Aombb

Here Aombb is the surface area of the same bounding box as calculated for the volume density (OMBB) feature.

3.1.21 Volume density (approximate enclosing ellipsoid) 6BDE

The eigenvectors and eigenvalues from PCA of the ROI voxel center point set Xc can be used to describe an ellipsoid approximating the point cloud 51, i.e. the approximate enclosing ellipsoid

(AEE). The volume of this ellipsoid is Vaee = 4π a b c/3, with a, b, and c being the lengths of the ellipsoid’s semi-principal axes, see Section 3.1.12. The volume density (AEE) is then:

3V F = morph.v.dens.aee 4πabc CHAPTER 3. IMAGE FEATURES 31

data value tol. consensus

dig. phantom 1.17 0.01 moderate config. A 1.29 0.01 strong config. B 1.29 0.01 strong config. C 1.29 0.01 moderate config. D 1.29 0.01 moderate config. E 1.29 0.01 strong

Table 3.20 — Reference values for the volume density (AEE) feature.

3.1.22 Area density (approximate enclosing ellipsoid) RDD2

The surface area of an ellipsoid can generally not be evaluated in an elementary form. However, it is possible to approximate the surface using an inﬁnite series. We use the same semi-principal axes as for the volume density (AEE) feature and deﬁne:

∞ ν X (α β) α2 + β2 A (a, b, c) = 4π a b P aee 1 − 4ν2 ν 2αβ ν=0

p 2 2 p 2 2 Here α = 1 − b /a and β = 1 − c /a are eccentricities of the ellipsoid and Pν is the Legendre polynomial function for degree ν. The Legendre polynomial series, though inﬁnite, converges, and approximation may be stopped early when the incremental gains in precision become limited. By default, we stop the series after ν = 20. The area density (AEE) is then approximated as:

A Fmorph.a.dens.aee = Aaee

data value tol. consensus

dig. phantom 1.36 0.01 moderate config. A 1.71 0.01 moderate config. B 1.62 0.01 moderate config. C 1.62 0.01 moderate config. D 1.62 0.01 moderate config. E 1.62 0.01 strong

Table 3.21 — Reference values for the area density (AEE) feature.

3.1.23 Volume density (minimum volume enclosing ellipsoid) SWZ1

Note: This feature currently has no reference values and should not be used. The minimum volume enclosing ellipsoid (MVEE), unlike the approximate enclosing ellipsoid, is the smallest ellipsoid that encloses the ROI. Direct computation of the MVEE is usually unfeasible, and is therefore approximated. Various approximation algorithms have been described, e.g. 2,72, which are usually elaborations on Khachiyan’s barycentric coordinate descent method 40.

The MVEE encloses the ROI mesh vertex set Xvx, and by deﬁnition Xvx,convex as well. Use of the convex mesh set Xvx,convex is recommended due to its sparsity compared to the full vertex CHAPTER 3. IMAGE FEATURES 32

set. The volume of the MVEE is deﬁned by its semi-axes lengths Vmvee = 4π a b c/3. Then:

V Fmorph.v.dens.mvee = Vmvee

For Khachiyan’s barycentric coordinate descent-based methods we use a default tolerance τ = 0.001 as stopping criterion.

3.1.24 Area density (minimum volume enclosing ellipsoid) BRI8

Note: This feature currently has no reference values and should not be used. The surface area of an ellipsoid does not have a general elementary form, but should be approximated as noted in Section 3.1.22. Let the approximated surface area of the MVEE be Amvee . Then: A Fmorph.a.dens.mvee = Amvee

3.1.25 Volume density (convex hull) R3ER

The convex hull encloses ROI mesh vertex set Xvx and consists of the vertex set Xvx,convex and corresponding faces, see section 3.1.11. The volume of the ROI mesh convex hull set Vconvex is computed in the same way as that of the volume (mesh) feature (3.1.1). The volume density can then be calculated as follows: V Fmorph.v.dens.conv.hull = Vconvex

This feature is also called solidity 27,65.

data value tol. consensus

dig. phantom 0.961 0.006 strong config. A 0.827 0.001 moderate config. B 0.829 0.001 moderate config. C 0.834 0.002 moderate config. D 0.834 0.002 moderate config. E 0.834 0.002 moderate

Table 3.22 — Reference values for the volume density (convex hull) feature.

3.1.26 Area density (convex hull) 7T7F

The area of the convex hull Aconvex is the sum of the areas of the faces of the convex hull, and is computed in the same way as the surface area (mesh) feature (section 3.1.3). The convex hull is identical to the one used in the volume density (convex hull) feature. Then:

A Fmorph.a.dens.conv.hull = Aconvex CHAPTER 3. IMAGE FEATURES 33

data value tol. consensus

dig. phantom 1.03 0.01 strong config. A 1.18 0.01 moderate config. B 1.12 0.01 moderate config. C 1.13 0.01 moderate config. D 1.13 0.01 moderate config. E 1.13 0.01 moderate

Table 3.23 — Reference values for the area density (convex hull) feature.

3.1.27 Integrated intensity 99N0

Integrated intensity is the average intensity in the ROI, multiplied by the volume. In the context of 18F-FDG-PET, this feature is often called total lesion glycolysis 75. Thus:

Nv,gl 1 X Fmorph.integ.int = V Xgl,k Nv,gl k=1

Nv,gl is the number of voxels in the ROI intensity mask.

data value tol. consensus

dig. phantom 1.2 × 103 10 moderate config. A 4.81 × 106 3.2 × 105 strong config. B 4.12 × 106 3.2 × 105 strong config. C −1.8 × 107 1.4 × 106 strong config. D −8.64 × 106 1.56 × 106 strong config. E −8.31 × 106 1.6 × 106 strong

Table 3.24 — Reference values for the integrated intensity feature.

3.1.28 Moran’s I index N365

Moran’s I index is an indicator of spatial autocorrelation 21,52. It is deﬁned as:

PNv,gl PNv,gl w (X − µ)(X − µ) Nv,gl k1=1 k2=1 k1k2 gl,k1 gl,k2 Fmorph.moran.i = , k1 6= k2 PNv,gl PNv,gl w PNv,gl (X − µ)2 k1=1 k2=1 k1k2 k=1 gl,k

As before Nv,gl is the number of voxels in the ROI intensity mask, µ is the mean of Xgl and wk1k2 is a weight factor, equal to the inverse Euclidean distance between voxels k1 and k2 of the point set 20 Xc,gl of the ROI intensity mask . Values of Moran’s I close to 1.0, 0.0 and -1.0 indicate high spatial autocorrelation, no spatial autocorrelation and high spatial anti-autocorrelation, respectively. Note that for an ROI containing many voxels, calculating Moran’s I index may be computationally expensive due to O(n2) behaviour. Approximation by repeated subsampling of the ROI may be required to make the calculation tractable, at the cost of accuracy. CHAPTER 3. IMAGE FEATURES 34

data value tol. consensus

dig. phantom 0.0397 0.0003 strong config. A 0.0322 0.0002 moderate config. B 0.0329 0.0001 moderate config. C 0.0824 0.0003 moderate config. D 0.0622 0.0013 moderate config. E 0.0596 0.0014 moderate

Table 3.25 — Reference values for the Moran’s I index feature.

3.1.29 Geary’s C measure NPT7

Geary’s C measure assesses spatial autocorrelation, similar to Moran’s I index 21,31. In contrast to Moran’s I index, Geary’s C measure directly assesses intensity diﬀerences between voxels and is more sensitive to local spatial autocorrelation. This measure is deﬁned as:

PNv,gl PNv,gl w (X − X )2 Nv,gl − 1 k1=1 k2=1 k1k2 gl,k1 gl,k2 Fmorph.geary.c = , k1 6= k2 2 PNv,gl PNv,gl w PNv,gl (X − µ)2 k1=1 k2=1 k1k2 k=1 gl,k

As with Moran’s I, Nv,gl is the number of voxels in the ROI intensity mask, µ is the mean of Xgl and wk1k2 is a weight factor, equal to the inverse Euclidean distance between voxels k1 and k2 of 20 the ROI voxel point set Xc,gl . Just as Moran’s I, Geary’s C measure exhibits O(n2) behaviour and an approximation scheme may be required to make calculation feasible for large ROIs.

data value tol. consensus

dig. phantom 0.974 0.006 strong config. A 0.863 0.001 moderate config. B 0.862 0.001 moderate config. C 0.846 0.001 moderate config. D 0.851 0.001 moderate config. E 0.853 0.001 moderate

Table 3.26 — Reference values for the Geary’s C measure feature. CHAPTER 3. IMAGE FEATURES 35

3.2 Local intensity features 9ST6

Voxel intensities within a deﬁned neighbourhood around a center voxel are used to compute local intensity features. Unlike many other feature sets, local features do not draw solely on intensities within the ROI. While only voxels within the ROI intensity map can be used as a center voxel, the local neighbourhood draws upon all voxels regardless of being in an ROI.

Aggregating features By deﬁnition, local intensity features are calculated in 3D (DHQ4), and not per slice.

3.2.1 Local intensity peak VJGA

The local intensity peak was originally devised for reducing variance in determining standardised uptake values 85. It is deﬁned as the mean intensity in a 1 cm3 spherical volume (in world coordinates), which is centered on the voxel with the maximum intensity level in the ROI intensity mask 29.

To calculate Floc.peak.local , we ﬁrst select all the voxels with centers within a radius r = 3 1/3 4π ≈ 0.62 cm of the center of the maximum intensity voxel. Subsequently, the mean intensity of the selected voxels, including the center voxel, are calculated. In case the maximum intensity is found in multiple voxels within the ROI, local intensity peak is calculated for each of these voxels, and the highest local intensity peak is chosen.

data value tol. consensus

dig. phantom 2.6 — strong config. A −277 10 moderate config. B 178 10 moderate config. C 169 10 moderate config. D 201 10 strong config. E 181 13 moderate

Table 3.27 — Reference values for the local intensity peak feature.

3.2.2 Global intensity peak 0F91

29 The global intensity peak feature Floc.peak.global is similar to the local intensity peak . However, instead of calculating the mean intensity for the voxel(s) with the maximum intensity, the mean intensity is calculated within a 1 cm3 neighbourhood for every voxel in the ROI intensity mask. The highest intensity peak value is then selected. Calculation of the global intensity peak feature may be accelerated by construction and application of an appropriate spatial spherical mean convolution filter, due to the convolution theorem. In this case one would first construct an empty 3D filter that will fit a 1 cm3 sphere. Within this context, the filter voxels may be represented by a point set, akin to Xc in section 3.1. Euclidean distances in world spacing between the central voxel of the filter and every remaining voxel are 3 1/3 computed. If this distance lies within radius r = 4π ≈ 0.62 the corresponding voxel receives a label 1, and 0 otherwise. Subsequent summation of the voxel labels yields Ns, the number of voxels within the 1 cm3 sphere. The filter then becomes a spherical mean filter by dividing the labels by Ns. CHAPTER 3. IMAGE FEATURES 36

data value tol. consensus

dig. phantom 3.1 — strong config. A 189 5 moderate config. B 178 5 moderate config. C 180 5 moderate config. D 201 5 moderate config. E 181 5 moderate

Table 3.28 — Reference values for the global intensity peak feature. CHAPTER 3. IMAGE FEATURES 37

3.3 Intensity-based statistical features UHIW

The intensity-based statistical features describe how intensities within the region of interest (ROI) are distributed. The features in this set do not require discretisation, and may be used to describe a continuous intensity distribution. Intensity-based statistical features are not meaningful if the intensity scale is arbitrary.

The set of intensities of the Nv voxels included in the ROI intensity mask is denoted as Xgl =

{Xgl,1,Xgl,2,...,Xgl,Nv }.

Aggregating features We recommend calculating intensity-based statistical features using the 3D volume (DHQ4). An approach that computes intensity-based statistical features per slice and subsequently averages them (3IDG) is not recommended.

3.3.1 Mean intensity Q4LE

The mean intensity of Xgl is calculated as:

N 1 Xv Fstat.mean = Xgl,k Nv k=1

data value tol. consensus

dig. phantom 2.15 — very strong config. A 13.4 1.1 very strong config. B 11.5 1.1 strong config. C −49 2.9 very strong config. D −23.5 3.9 strong config. E −22.6 4.1 strong

Table 3.29 — Reference values for the mean feature.

3.3.2 Intensity variance ECT3

The intensity variance of Xgl is deﬁned as:

Nv 1 X 2 Fstat.var = (Xgl,k − µ) Nv k=1

Note that we do not apply a bias correction when computing the variance. CHAPTER 3. IMAGE FEATURES 38

data value tol. consensus

dig. phantom 3.05 — very strong config. A 1.42 × 104 400 very strong config. B 1.44 × 104 400 strong config. C 5.06 × 104 1.4 × 103 very strong config. D 3.28 × 104 2.1 × 103 strong config. E 3.51 × 104 2.2 × 103 strong

Table 3.30 — Reference values for the variance feature.

3.3.3 Intensity skewness KE2A

The skewness of the intensity distribution of Xgl is deﬁned as:

1 PNv 3 (Xgl,k − µ) F = Nv k=1 stat.skew 3/2 1 PNv 2 (Xgl,k − µ) Nv k=1

Here µ = Fstat.mean . If the intensity variance Fstat.var = 0, Fstat.skew = 0.

data value tol. consensus

dig. phantom 1.08 — very strong config. A −2.47 0.05 very strong config. B −2.49 0.05 strong config. C −2.14 0.05 very strong config. D −2.28 0.06 strong config. E −2.3 0.07 strong

Table 3.31 — Reference values for the skewness feature.

3.3.4 (Excess) intensity kurtosis IPH6

Kurtosis, or technically excess kurtosis, is a measure of peakedness in the intensity distribution

Xgl:

1 PNv 4 (Xgl,k − µ) F = Nv k=1 − 3 stat.kurt 2 1 PNv 2 (Xgl,k − µ) Nv k=1

Here µ = Fstat.mean . Note that kurtosis is corrected by a Fisher correction of -3 to center it on 0 for normal distributions. If the intensity variance Fstat.var = 0, Fstat.kurt = 0. CHAPTER 3. IMAGE FEATURES 39

data value tol. consensus

dig. phantom −0.355 — very strong config. A 5.96 0.24 very strong config. B 5.93 0.24 strong config. C 3.53 0.23 very strong config. D 4.35 0.32 strong config. E 4.44 0.33 strong

Table 3.32 — Reference values for the (excess) kurtosis feature.

3.3.5 Median intensity Y12H

The median intensity Fstat.median is the sample median of Xgl.

data value tol. consensus

dig. phantom 1 — very strong config. A 46 0.3 very strong config. B 45 0.3 strong config. C 40 0.4 very strong config. D 42 0.4 strong config. E 43 0.5 strong

Table 3.33 — Reference values for the median feature.

3.3.6 Minimum intensity 1GSF

The minimum intensity is equal to the lowest intensity present in Xgl, i.e:

Fstat.min = min(Xgl)

data value tol. consensus

dig. phantom 1 — very strong config. A −500 — very strong config. B −500 — strong config. C −939 4 very strong config. D −724 12 strong config. E −743 13 strong

Table 3.34 — Reference values for the minimum feature.

3.3.7 10th intensity percentile QG58

th P10 is the 10 percentile of Xgl. P10 is a more robust alternative to the minimum intensity. CHAPTER 3. IMAGE FEATURES 40

data value tol. consensus

dig. phantom 1 — very strong config. A −129 8 very strong config. B −136 8 strong config. C −424 14 very strong config. D −304 20 strong config. E −310 21 strong

Table 3.35 — Reference values for the 10th percentile feature.

3.3.8 90th intensity percentile 8DWT

th P90 is the 90 percentile of Xgl. P90 is a more robust alternative to the maximum intensity.

data value tol. consensus

dig. phantom 4 — very strong config. A 95 — strong config. B 91 — strong config. C 86 0.1 strong config. D 86 0.1 strong config. E 93 0.2 strong

Table 3.36 — Reference values for the 90th percentile feature.

Note that the 90th intensity percentile obtained for the digital phantom may diﬀer from the above reference value depending on the software implementation used to compute it. For example, some implementations were found to produce a value of 4.2 instead of 4.

3.3.9 Maximum intensity 84IY

The maximum intensity is equal to the highest intensity present in Xgl, i.e:

Fstat.max = max(Xgl)

data value tol. consensus

dig. phantom 6 — very strong config. A 377 9 very strong config. B 391 9 strong config. C 393 10 very strong config. D 521 22 strong config. E 345 9 strong

Table 3.37 — Reference values for the maximum feature.

3.3.10 Intensity interquartile range SALO

The interquartile range (IQR) of Xgl is deﬁned as:

Fstat.iqr = P75 − P25 CHAPTER 3. IMAGE FEATURES 41

th th P25 and P75 are the 25 and 75 percentiles of Xgl, respectively.

data value tol. consensus

dig. phantom 3 — very strong config. A 56 0.5 very strong config. B 52 0.5 strong config. C 67 4.9 very strong config. D 57 4.1 strong config. E 62 3.5 strong

Table 3.38 — Reference values for the interquartile range feature.

3.3.11 Intensity range 2OJQ

The intensity range is deﬁned as:

Fstat.range = max(Xgl) − min(Xgl)

data value tol. consensus

dig. phantom 5 — very strong config. A 877 9 very strong config. B 891 9 strong config. C 1.33 × 103 20 very strong config. D 1.24 × 103 40 strong config. E 1.09 × 103 30 strong

Table 3.39 — Reference values for the range feature.

3.3.12 Intensity-based mean absolute deviation 4FUA

Mean absolute deviation is a measure of dispersion from the mean of Xgl:

N 1 Xv Fstat.mad = |Xgl,k − µ| Nv k=1

Here µ = Fstat.mean .

data value tol. consensus

dig. phantom 1.55 — very strong config. A 73.6 1.4 very strong config. B 74.4 1.4 strong config. C 158 4 very strong config. D 123 6 strong config. E 125 6 strong

Table 3.40 — Reference values for the mean absolute deviation feature. CHAPTER 3. IMAGE FEATURES 42

3.3.13 Intensity-based robust mean absolute deviation 1128

The intensity-based mean absolute deviation feature may be inﬂuenced by outliers. To increase robustness, the set of intensities can be restricted to those which lie closer to the center of the distribution. Let

Xgl,10−90 = {x ∈ Xgl|P10 (Xgl) ≤ x ≤ P90 (Xgl)}

Then Xgl,10−90 is the set of Nv,10−90 ≤ Nv voxels in Xgl whose intensities fall in the interval th th bounded by the 10 and 90 percentiles of Xgl. The robust mean absolute deviation is then:

Nv,10−90 1 X Fstat.rmad = Xgl,10−90,k − Xgl,10−90 Nv,10−90 k=1

Xgl,10−90 denotes the sample mean of Xgl,10−90.

data value tol. consensus

dig. phantom 1.11 — very strong config. A 27.7 0.8 very strong config. B 27.3 0.8 strong config. C 66.8 3.5 very strong config. D 46.8 3.6 strong config. E 46.5 3.7 strong

Table 3.41 — Reference values for the robust mean absolute deviation feature.

3.3.14 Intensity-based median absolute deviation N72L

Median absolute deviation is similar in concept to the intensity-based mean absolute deviation, but measures dispersion from the median intensity instead of the mean intensity. Thus:

N 1 Xv Fstat.medad = |Xgl,k − M| Nv k=1

Here, median M = Fstat.median .

data value tol. consensus

dig. phantom 1.15 — very strong config. A 64.3 1 strong config. B 63.8 1 strong config. C 119 4 strong config. D 94.7 3.8 strong config. E 97.9 3.9 strong

Table 3.42 — Reference values for the median absolute deviation feature. CHAPTER 3. IMAGE FEATURES 43

3.3.15 Intensity-based coeﬃcient of variation 7TET

The coeﬃcient of variation measures the dispersion of Xgl. It is deﬁned as: σ F = stat.cov µ

1/2 Here σ = Fstat.var and µ = Fstat.mean are the standard deviation and mean of the intensity distribution, respectively.

data value tol. consensus

dig. phantom 0.812 — very strong config. A 8.9 4.98 strong config. B 10.4 5.2 strong config. C −4.59 0.29 strong config. D −7.7 1.01 strong config. E −8.28 0.95 strong

Table 3.43 — Reference values for the coeﬃcient of variation feature.

3.3.16 Intensity-based quartile coeﬃcient of dispersion 9S40

The quartile coefficient of dispersion is a more robust alternative to the intensity-based coefficient of variance. It is defined as:

P75 − P25 Fstat.qcod = P75 + P25

th th P25 and P75 are the 25 and 75 percentile of Xgl, respectively.

data value tol. consensus

dig. phantom 0.6 — very strong config. A 0.636 0.008 strong config. B 0.591 0.008 strong config. C 1.03 0.4 strong config. D 0.74 0.011 strong config. E 0.795 0.337 strong

Table 3.44 — Reference values for the quartile coeﬃcient of dispersion feature.

3.3.17 Intensity-based energy N8CA

1 The energy of Xgl is deﬁned as:

Nv X 2 Fstat.energy = Xgl,k k=1 CHAPTER 3. IMAGE FEATURES 44

data value tol. consensus

dig. phantom 567 — very strong config. A 1.65 × 109 2 × 107 very strong config. B 3.98 × 108 1.1 × 107 strong config. C 2.44 × 109 1.2 × 108 strong config. D 1.48 × 109 1.4 × 108 strong config. E 1.58 × 109 1.4 × 108 strong

Table 3.45 — Reference values for the energy feature.

3.3.18 Root mean square intensity 5ZWQ

1 The root mean square intensity feature , which is also called the quadratic mean, of Xgl is deﬁned as: s PNv 2 k=1 Xgl,k Fstat.rms = Nv

data value tol. consensus

dig. phantom 2.77 — very strong config. A 120 2 very strong config. B 121 2 strong config. C 230 4 strong config. D 183 7 strong config. E 189 7 strong

Table 3.46 — Reference values for the root mean square feature. CHAPTER 3. IMAGE FEATURES 45

3.4 Intensity histogram features ZVCW

An intensity histogram is generated by discretising the original intensity distribution Xgl into intensity bins. Approaches to discretisation are described in Section 2.7.

Let Xd = {Xd,1,Xd,2,...,Xd,Nv } be the set of Ng discretised intensities of the Nv voxels in the ROI intensity mask. Let H = n1, n2, . . . , nNg be the histogram with frequency count ni of each discretised intensity i in Xd. The occurrence probability pi for each discretised intensity i is then approximated as pi = ni/Nv.

Aggregating features We recommend calculating intensity histogram features using the 3D volume (DHQ4). An approach that computes features per slice and subsequently averages (3IDG) is not recommended.

3.4.1 Mean discretised intensity X6K6

1 The mean of Xd is calculated as:

N 1 Xv Fih.mean = Xd,k Nv k=1

An equivalent deﬁnition is:

Ng X Fih.mean = i pi i=1

data value tol. consensus

dig. phantom 2.15 — very strong config. A 21.1 0.1 strong config. B 18.9 0.3 strong config. C 38.6 0.2 strong config. D 18.5 0.5 strong config. E 21.7 0.3 strong

Table 3.47 — Reference values for the mean feature.

3.4.2 Discretised intensity variance CH89

1 The variance of Xd is deﬁned as:

Nv 1 X 2 Fih.var = (Xd,k − µ) Nv k=1

Here µ = Fih.mean . This deﬁnition is equivalent to:

Ng X 2 Fih.var = (i − µ) pi i=1

Note that no bias-correction is applied when computing the variance. CHAPTER 3. IMAGE FEATURES 46

data value tol. consensus

dig. phantom 3.05 — strong config. A 22.8 0.6 strong config. B 18.7 0.2 strong config. C 81.1 2.1 strong config. D 21.7 0.4 strong config. E 30.4 0.8 strong

Table 3.48 — Reference values for the variance feature.

3.4.3 Discretised intensity skewness 88K1

1 The skewness of Xd is deﬁned as:

1 PNv 3 (Xd,k − µ) F = Nv k=1 ih.skew 3/2 1 PNv 2 (Xd,k − µ) Nv k=1

Here µ = Fih.mean . This deﬁnition is equivalent to:

PNg (i − µ)3 p F = i=1 i ih.skew 3/2 PNg 2 i=1 (i − µ) pi

If the discretised intensity variance Fih.var = 0, Fih.skew = 0.

data value tol. consensus

dig. phantom 1.08 — very strong config. A −2.46 0.05 strong config. B −2.47 0.05 strong config. C −2.14 0.05 strong config. D −2.27 0.06 strong config. E −2.29 0.07 strong

Table 3.49 — Reference values for the skewness feature.

3.4.4 (Excess) discretised intensity kurtosis C3I7

1 Kurtosis , or technically excess kurtosis, measures the peakedness of the Xd distribution:

1 PNv 4 (Xd,k − µ) F = Nv k=1 − 3 ih.kurt 2 1 PNv 2 (Xd,k − µ) Nv k=1

Here µ = Fih.mean . An alternative, but equivalent, deﬁnition is:

PNg (i − µ)4 p F = i=1 i − 3 ih.kurt 2 PNg 2 i=1 (i − µ) pi

Note that kurtosis is corrected by a Fisher correction of -3 to center kurtosis on 0 for normal distributions. If the discretised intensity variance Fih.var = 0, Fih.kurt = 0. CHAPTER 3. IMAGE FEATURES 47

data value tol. consensus

dig. phantom −0.355 — very strong config. A 5.9 0.24 strong config. B 5.84 0.24 strong config. C 3.52 0.23 strong config. D 4.31 0.32 strong config. E 4.4 0.33 strong

Table 3.50 — Reference values for the (excess) kurtosis feature.

3.4.5 Median discretised intensity WIFQ

1 The median Fih.median is the sample median of Xd .

data value tol. consensus

dig. phantom 1 — very strong config. A 22 — strong config. B 20 0.3 strong config. C 42 — strong config. D 20 0.5 strong config. E 24 0.2 strong

Table 3.51 — Reference values for the median feature.

3.4.6 Minimum discretised intensity 1PR8

1 The minimum discretised intensity is equal to the lowest discretised intensity present in Xd, i.e.:

Fih.min = min(Xd)

For fixed bin number discretisation Fih.min = 1 by definition, but Fih.min > 1 is possible for fixed bin size discretisation.

data value tol. consensus

dig. phantom 1 — very strong config. A 1 — strong config. B 1 — strong config. C 3 0.16 strong config. D 1 — strong config. E 1 — strong

Table 3.52 — Reference values for the minimum feature.

3.4.7 10th discretised intensity percentile GPMT

th P10 is the 10 percentile of Xd. CHAPTER 3. IMAGE FEATURES 48

data value tol. consensus

dig. phantom 1 — very strong config. A 15 0.4 strong config. B 14 0.5 strong config. C 24 0.7 strong config. D 11 0.7 strong config. E 13 0.7 strong

Table 3.53 — Reference values for the 10th percentile feature.

3.4.8 90th discretised intensity percentile OZ0C

th P90 is the 90 percentile of Xd and is deﬁned as Fih.P90 .

data value tol. consensus

dig. phantom 4 — strong config. A 24 — strong config. B 22 0.3 strong config. C 44 — strong config. D 21 0.5 strong config. E 25 0.2 strong

Table 3.54 — Reference values for the 90th percentile feature.

Note that the 90th discretised intensity percentile obtained for the digital phantom may diﬀer from the above reference value depending on the software implementation used to compute it. For example, some implementations were found to produce a value of 4.2 instead of 4 for this feature.

3.4.9 Maximum discretised intensity 3NCY

1 The maximum discretised intensity is equal to the highest discretised intensity present in Xd, i.e.:

Fih.max = max(Xd)

By deﬁnition, Fih.max = Ng.

data value tol. consensus

dig. phantom 6 — very strong config. A 36 0.4 strong config. B 32 — strong config. C 56 0.5 strong config. D 32 — strong config. E 32 — strong

Table 3.55 — Reference values for the maximum feature. CHAPTER 3. IMAGE FEATURES 49

3.4.10 Intensity histogram mode AMMC

The mode of Xd Fih.mode is the most common discretised intensity present, i.e. the value i for with the highest count ni. The mode may not be uniquely deﬁned. When the highest count is found in multiple bins, the value i of the bin closest to the mean discretised intensity is chosen as intensity histogram mode. In pathological cases with two such bins equidistant to the mean, the bin to the left of the mean is selected.

data value tol. consensus

dig. phantom 1 — very strong config. A 23 — strong config. B 20 0.3 strong config. C 43 0.1 strong config. D 20 0.4 strong config. E 24 0.1 strong

Table 3.56 — Reference values for the mode feature.

3.4.11 Discretised intensity interquartile range WR0O

The interquartile range (IQR) of Xd is deﬁned as:

Fih.iqr = P75 − P25

th th P25 and P75 are the 25 and 75 percentile of Xd, respectively.

data value tol. consensus

dig. phantom 3 — very strong config. A 2 — strong config. B 2 — strong config. C 3 0.21 strong config. D 2 0.06 strong config. E 1 0.06 strong

Table 3.57 — Reference values for the interquartile range feature.

3.4.12 Discretised intensity range 5Z3W

The discretised intensity range 1 is deﬁned as:

Fih.range = max(Xd) − min(Xd)

For ﬁxed bin number discretisation, the discretised intensity range equals Ng by deﬁnition. CHAPTER 3. IMAGE FEATURES 50

data value tol. consensus

dig. phantom 5 — very strong config. A 35 0.4 strong config. B 31 — strong config. C 53 0.6 strong config. D 31 — strong config. E 31 — strong

Table 3.58 — Reference values for the range feature.

3.4.13 Intensity histogram mean absolute deviation D2ZX

1 The mean absolute deviation is a measure of dispersion from the mean of Xd:

N 1 Xv F = |X − µ| ih.mad N d,i v i=1

Here µ = Fih.mean .

data value tol. consensus

dig. phantom 1.55 — very strong config. A 2.94 0.06 strong config. B 2.67 0.03 strong config. C 6.32 0.15 strong config. D 3.15 0.05 strong config. E 3.69 0.1 strong

Table 3.59 — Reference values for the mean absolute deviation feature.

3.4.14 Intensity histogram robust mean absolute deviation WRZB

Intensity histogram mean absolute deviation may be aﬀected by outliers. To increase robustness, the set of discretised intensities under consideration can be restricted to those which are closer to the center of the distribution. Let

Xd,10−90 = {x ∈ Xd|P10 (Xd) ≤ x ≤ P90 (Xd)}

In short, Xd,10−90 is the set of Nv,10−90 ≤ Nv voxels in Xd whose discretised intensities fall in the th th interval bounded by the 10 and 90 percentiles of Xd. The robust mean absolute deviation is then:

Nv,10−90 1 X Fih.rmad = Xd,10−90,k − Xd,10−90 Nv,10−90 k=1

Xd,10−90 denotes the sample mean of Xd,10−90. CHAPTER 3. IMAGE FEATURES 51

data value tol. consensus

dig. phantom 1.11 — very strong config. A 1.18 0.04 strong config. B 1.03 0.03 moderate config. C 2.59 0.14 strong config. D 1.33 0.06 strong config. E 1.46 0.09 moderate

Table 3.60 — Reference values for the robust mean absolute deviation feature.

3.4.15 Intensity histogram median absolute deviation 4RNL

Histogram median absolute deviation is conceptually similar to histogram mean absolute deviation, but measures dispersion from the median instead of mean. Thus:

N 1 Xv Fih.medad = |Xd,k − M| Nv k=1

Here, median M = Fih.median .

data value tol. consensus

dig. phantom 1.15 — very strong config. A 2.58 0.05 strong config. B 2.28 0.02 strong config. C 4.75 0.12 strong config. D 2.41 0.04 strong config. E 2.89 0.07 strong

Table 3.61 — Reference values for the median absolute deviation feature.

3.4.16 Intensity histogram coeﬃcient of variation CWYJ

The coeﬃcient of variation measures the dispersion of the discretised intensity distribution. It is deﬁned as: σ F = ih.cov µ

1/2 Here σ = Fih.var and µ = Fih.mean are the standard deviation and mean of the discretised intensity distribution, respectively.

data value tol. consensus

dig. phantom 0.812 — very strong config. A 0.227 0.004 strong config. B 0.229 0.004 strong config. C 0.234 0.005 strong config. D 0.252 0.006 strong config. E 0.254 0.006 strong

Table 3.62 — Reference values for the coeﬃcient of variation feature. CHAPTER 3. IMAGE FEATURES 52

3.4.17 Intensity histogram quartile coeﬃcient of dispersion SLWD

The quartile coefficient of dispersion is a more robust alternative to the intensity histogram coefficient of variance. It is defined as:

P75 − P25 Fih.qcod = P75 + P25

th th P25 and P75 are the 25 and 75 percentile of Xd, respectively.

data value tol. consensus

dig. phantom 0.6 — very strong config. A 0.0455 — strong config. B 0.05 0.0005 strong config. C 0.0361 0.0027 strong config. D 0.05 0.0021 strong config. E 0.0213 0.0015 strong

Table 3.63 — Reference values for the quartile coeﬃcient of dispersion feature.

3.4.18 Discretised intensity entropy TLU2

Entropy 1 is an information-theoretic concept that gives a metric for the information contained within Xd. The particular metric used is Shannon entropy, which is deﬁned as:

Ng X Fih.entropy = − pi log2 pi i=1

Note that entropy can only be meaningfully deﬁned for discretised intensities as it will tend to

− log2 Nv for continuous intensity distributions.

data value tol. consensus

dig. phantom 1.27 — very strong config. A 3.36 0.03 very strong config. B 3.16 0.01 strong config. C 3.73 0.04 strong config. D 2.94 0.01 strong config. E 3.22 0.02 strong

Table 3.64 — Reference values for the entropy feature.

3.4.19 Discretised intensity uniformity BJ5W

1 Uniformity of Xd is deﬁned as:

Ng X 2 Fih.uniformity = pi i=1

For histograms where most intensities are contained in a single bin, uniformity approaches 1.

The lower bound is 1/Ng. CHAPTER 3. IMAGE FEATURES 53

Note that this feature is sometimes referred to as energy.

data value tol. consensus

dig. phantom 0.512 — very strong config. A 0.15 0.002 very strong config. B 0.174 0.001 strong config. C 0.14 0.003 strong config. D 0.229 0.003 strong config. E 0.184 0.001 strong

Table 3.65 — Reference values for the uniformity feature.

3.4.20 Maximum histogram gradient 12CE

The histogram gradient H0 of intensity histogram H can be calculated as:  n − n i = 1  2 1 0  Hi = (ni+1 − ni−1) /2 1 < i < Ng  nNg − nNg −1 i = Ng

Histogram H should be non-sparse, i.e. bins where ni = 0 should not be omitted. Ostensibly, the histogram gradient can be calculated in diﬀerent ways. The method above has the advantages of being easy to implement and leading to a gradient H0 with same size as H. This helps maintain a direct correspondence between the discretised intensities in H and the bins of H0. The maximum histogram gradient 80 is:

0 Fih.max.grad = max (H )

data value tol. consensus

dig. phantom 8 — very strong config. A 1.1 × 104 100 strong config. B 3.22 × 103 50 strong config. C 4.75 × 103 30 strong config. D 7.26 × 103 200 strong config. E 6.01 × 103 130 strong

Table 3.66 — Reference values for the maximum histogram gradient feature.

3.4.21 Maximum histogram gradient intensity 8E6O

80 The maximum histogram gradient intensity Fih.max.grad.gl is the discretised intensity corresponding to the maximum histogram gradient, i.e. the value i in H for which H0 is maximal. CHAPTER 3. IMAGE FEATURES 54

data value tol. consensus

dig. phantom 3 — strong config. A 21 — strong config. B 19 0.3 strong config. C 41 — strong config. D 19 0.4 strong config. E 23 0.2 moderate

Table 3.67 — Reference values for the maximum histogram gradient intensity feature.

3.4.22 Minimum histogram gradient VQB3

The minimum histogram gradient 80 is:

0 Fih.min.grad = min (H )

data value tol. consensus

dig. phantom −50 — very strong config. A −1.01 × 104 100 strong config. B −3.02 × 103 50 strong config. C −4.68 × 103 50 strong config. D −6.67 × 103 230 strong config. E −6.11 × 103 180 strong

Table 3.68 — Reference values for the minimum histogram gradient feature.

3.4.23 Minimum histogram gradient intensity RHQZ

80 The minimum histogram gradient intensity Fih.min.grad.gl is the discretised intensity corresponding to the minimum histogram gradient, i.e. the value i in H for which H0 is minimal.

data value tol. consensus

dig. phantom 1 — strong config. A 24 — strong config. B 22 0.3 strong config. C 44 — strong config. D 22 0.4 strong config. E 25 0.2 strong

Table 3.69 — Reference values for the minimum histogram gradient intensity feature. CHAPTER 3. IMAGE FEATURES 55

3.5 Intensity-volume histogram features P88C

The (cumulative) intensity-volume histogram (IVH) of the set Xgl of voxel intensities in the ROI intensity mask describes the relationship between discretised intensity i and the fraction of the volume containing at least intensity i, ν 27.

Depending on the imaging modality, the calculation of IVH features requires discretising Xgl to generate a new voxel set Xd,gl with discretised intensities. Moreover, the total range G of discretised intensities and the discretisation interval wd should be provided or determined. The total range G determines the range of discretised intensities to be included in the IVH, whereas the discretisation interval determines the intensity diﬀerence between adjacent discretised intensities in the IVH. Recommendations for discretisation parameters diﬀer depending on what type of data the image represents, and how it is represented. These recommendations are described below.

Discrete calibrated image intensities Some imaging modalities by default generate voxels with calibrated, discrete intensities – for example CT. In this case, the discretised ROI voxel set Xd,gl = Xgl, i.e. no discretisation required. If a re-segmentation range is provided (see Section 2.5), the total range G is equal to the re- segmentation range. In the case of a half-open re-segmentation range, the upper limit of the range is max(Xgl). When no range is provided, G = [min(Xgl), max(Xgl)]. The discretisation interval is wd = 1.

Continuous calibrated image intensities Imaging with calibrated, continuous intensities such as PET requires discretisation to determine the IVH, while preserving the quantitative intensity information. The use of a ﬁxed bin size discretisation method is thus recommended, see Section 2.7. This method requires a minimum intensity Xgl,min, a maximum intensity Xgl,max and the bin width wb. If a re-segmentation range is deﬁned (see Section 2.5), Xgl,min is set to the lower bound of the re-segmentation range and Xgl,max to the upper bound; otherwise Xgl,min = min(Xgl) and Xgl,max = max(Xgl) (i.e. the minimum and maximum intensities in the imaging volume prior to discretisation). The bin width wb is modality dependent, but should be small relative to the intensity range, e.g. 0.10 SUV for 18F-FDG-PET.

Next, ﬁxed bin size discretisation produces the voxel set Xd of bin numbers, which needs to be converted to bin centers in order to maintain a direct relationship with the original intensities. We thus replace bin numbers Xd with the intensity corresponding to the bin center:

Xd,gl = Xgl,min + (Xd − 0.5) wb

The total range is then G = [Xgl,min + 0.5wb,Xgl,max − 0.5wb]. In this case, the discretisation interval matches the bin width, i.e. wd = wb.

Arbitrary intensity units Some imaging modalities, such as many MRI sequences, produce arbitrary intensities. In such cases, a ﬁxed bin number discretisation method with Ng = 1000 bins is recommended, see Section 2.7. The discretisation bin width is wb = (Xgl,max − Xgl,min) /Ng, with Xgl,max = max (Xgl) and Xgl,min = min (Xgl), as re-segmentation ranges generally cannot be provided for non-calibrated intensities. The ﬁxed bin number discretisation produces the voxel set Xd ∈ {1, 2,...,Ng}. Because of the lack of calibration, Xd,gl = Xd, and consequentially the discretisation interval is wd = 1 and the total range is G = [1,Ng] CHAPTER 3. IMAGE FEATURES 56

i γ ν 1 0.0 1.000 2 0.2 0.324 3 0.4 0.324 4 0.6 0.311 5 0.8 0.095 6 1.0 0.095

Table 3.70 — Example intensity-volume histogram evaluated at discrete intensities i of the digital phantom. The total range G = [1, 6], with discretisation interval w = 1. Thus γ is the intensity fraction and ν is the corresponding volume fraction that contains intensity i or greater.

Calculating the IV histogram

We use Xd,gl to calculate fractional volumes and fractional intensities. As voxels for the same image stack generally all have the same dimensions, we may deﬁne fractional volume ν for discretised intensity i:

N 1 Xv νi = 1 − [Xd,gl,k < i] Nv k=1

Here [...] is an Iverson bracket, yielding 1 if the condition is true and 0 otherwise. In essence, we count the voxels containing a discretised intensity smaller than i, divide by the total number of voxels, and then subtract this volume fraction to ﬁnd νi. The intensity fraction γ for discretised intensity i in the range G is calculated as:

i − min (G) γ = i max (G) − min (G)

Note that intensity fractions are also calculated for discretised intensities that are absent in Xd,gl. For example intensities 2 and 5 are absent in the digital phantom (see Chapter 5), but are still evaluated to determine both the fractional volume and the intensity fraction. An example IVH for the digital phantom is shown in Table 3.70.

Aggregating features We recommend calculating intensity-volume histogram features using the 3D volume (DHQ4). Computing features per slice and subsequently averaging (3IDG) is not recommended.

3.5.1 Volume at intensity fraction BC2M

The volume at intensity fraction Vx is the largest volume fraction ν that has an intensity fraction γ of at least x%. This differs from conceptually similar dose-volume histograms used in radiotherapy planning, where V10 would indicate the volume fraction receiving at least 10 Gy planned dose. El 27 Naqa et al. defined both V10 and V90 as features. In the context of this work, these two features are defined as Fivh.V10 and Fivh.V90 , respectively. CHAPTER 3. IMAGE FEATURES 57

data value tol. consensus

dig. phantom 0.324 — very strong config. A 0.978 0.001 strong config. B 0.977 0.001 strong config. C 0.998 0.001 moderate config. D 0.972 0.003 strong config. E 0.975 0.002 strong

Table 3.71 — Reference values for the volume fraction at 10% intensity feature.

data value tol. consensus

dig. phantom 0.0946 — very strong config. A 6.98 × 10−5 1.03 × 10−5 strong config. B 7.31 × 10−5 1.03 × 10−5 strong config. C 0.000152 2 × 10−5 strong config. D 9 × 10−5 0.000415 strong config. E 0.000157 0.000248 strong

Table 3.72 — Reference values for the volume fraction at 90% intensity feature.

3.5.2 Intensity at volume fraction GBPN

The intensity at volume fraction Ix is the minimum discretised intensity i present in at most x% 27 of the volume. El Naqa et al. deﬁned both I10 and I90 as features. In the context of this work, these two features are deﬁned as Fivh.I10 and Fivh.I90 , respectively.

data value tol. consensus

dig. phantom 5 — very strong config. A 96 — strong config. B 92 — strong config. C 88.8 0.2 moderate config. D 87 0.1 strong config. E 770 5 moderate

Table 3.73 — Reference values for the intensity at 10% volume feature. CHAPTER 3. IMAGE FEATURES 58

data value tol. consensus

dig. phantom 2 — very strong config. A −128 8 strong config. B −135 8 strong config. C −421 14 strong config. D −303 20 strong config. E 399 17 moderate

Table 3.74 — Reference values for the intensity at 90% volume feature.

3.5.3 Volume fraction diﬀerence between intensity fractions DDTU

This feature is the difference between the volume fractions at two different intensity fractions, e.g. 27 V10 − V90 . In the context of this work, this feature is defined as Fivh.V10minusV90 .

data value tol. consensus

dig. phantom 0.23 — very strong config. A 0.978 0.001 strong config. B 0.977 0.001 strong config. C 0.997 0.001 strong config. D 0.971 0.001 strong config. E 0.974 0.001 strong

Table 3.75 — Reference values for the volume fraction diﬀerence between 10% and 90% intensity feature.

3.5.4 Intensity fraction diﬀerence between volume fractions CNV2

This feature is the difference between discretised intensities at two different fractional volumes, 27 e.g. I10 − I90 . In the context of this work, this feature is defined as Fivh.I10minusI90 .

data value tol. consensus

dig. phantom 3 — very strong config. A 224 8 strong config. B 227 8 strong config. C 510 14 strong config. D 390 20 strong config. E 371 13 moderate

Table 3.76 — Reference values for the intensity diﬀerence between 10% and 90% volume feature.

3.5.5 Area under the IVH curve 9CMM

Note: This feature currently has no reference values and should not be used.

82 The area under the IVH curve Fivh.auc was deﬁned by van Velden et al. . The area under the IVH curve can be approximated by calculating the Riemann sum using the trapezoidal rule. Note that if there is only one discretised intensity in the ROI, we deﬁne the area under the IVH curve

Fivh.auc = 0. CHAPTER 3. IMAGE FEATURES 59

3.6 Grey level co-occurrence based features LFYI

In image analysis, texture is one of the deﬁning sets of features. Texture features were originally designed to assess surface texture in 2D images. Texture analysis is however not restricted to 2D slices and can be extended to 3D objects. Image intensities are generally discretised before calculation of texture features, see Section 2.7. The grey level co-occurrence matrix (GLCM) is a matrix that expresses how combinations of discretised intensities (grey levels) of neighbouring pixels, or voxels in a 3D volume, are distributed along one of the image directions. Generally, the neighbourhood for GLCM is a 26-connected neighbourhood in 3D and a 8-connected neighbourhood in 2D. Thus, in 3D there are 13 unique direction vectors within the neighbourhood for Chebyshev distance δ = 1, i.e. (0, 0, 1), (0, 1, 0), (1, 0, 0), (0, 1, 1), (0, 1, −1), (1, 0, 1), (1, 0, −1), (1, 1, 0), (1, −1, 0), (1, 1, 1), (1, 1, −1), (1, −1, 1) and (1, −1, −1), whereas in 2D the direction vectors are (1, 0, 0), (1, 1, 0), (0, 1, 0) and (−1, 1, 0).

A GLCM is calculated for each direction vector, as follows. Let Mm be the Ng × Ng grey level co-occurrence matrix, with Ng the number of discretised grey levels present in the ROI intensity mask, and m the particular direction vector. Element (i, j) of the GLCM contains the frequency at which combinations of discretised grey levels i and j occur in neighbouring voxels along direction m = m and along direction m = −m. Then, M = M + M = M + MT 36. As a + − m m+ m− m+ m+ consequence the GLCM matrix Mm is symmetric. An example of the calculation of a GLCM is shown in Table 3.77. Corresponding grey level co-occurrence matrices for each direction are shown in Table 3.78.

j j 1 2 2 3 0 3 0 0 0 0 0 2 1 2 3 3 0 1 3 1 3 1 0 1 i i 4 2 4 1 0 0 1 0 0 3 1 0 4 1 2 3 2 1 0 0 0 1 0 0

(a) Grey levels (b) Mm+=→ (c) Mm−=← Table 3.77 — Grey levels (a) and corresponding grey level co-occurrence matrices for the 0◦ (b) ◦ and 180 directions (c). In vector notation these directions are m+ = (1, 0) and m− = (−1, 0). To calculate the symmetrical co-occurrence matrix Mm both matrices are summed by element.

GLCM features rely on the probability distribution for the elements of the GLCM. Let us consider Mm=(1,0) from the example, as shown in Table 3.79. We derive a probability distribution for grey level co-occurrences, Pm, by normalising Mm by the sum of its elements. Each element pij of Pm is then the joint probability of grey levels i and j occurring in neighbouring voxels PNg PNg along direction m. Then pi. = j=1 pij is the row marginal probability, and p.j = i=1 pij is the column marginal probability. As Pm is by deﬁnition symmetric, pi. = p.j. Furthermore, let us 36,74 consider diagonal and cross-diagonal probabilities pi−j and pi+j :

Ng Ng X X pi−j,k = pij [k = |i − j|] k = 0,...,Ng − 1 i=1 j=1

Ng Ng X X pi+j,k = pij [k = i + j] k = 2,..., 2Ng i=1 j=1

Here, [...] is an Iverson bracket, which equals 1 when the condition within the brackets is true and 0 otherwise. In eﬀect we select only combinations of elements (i, j) for which the condition holds. CHAPTER 3. IMAGE FEATURES 60

j j 0 3 0 2 0 2 0 1 3 2 3 2 2 2 1 2 i i 0 3 2 0 0 1 2 1 2 2 0 0 1 2 1 0

(a) Mm=→ (b) Mm=%

j j 2 1 2 1 0 2 1 1 1 4 1 1 2 2 2 1 i i 2 1 2 1 1 2 0 1 1 1 1 2 1 1 1 0

(c) Mm=↑ (d) Mm=- Table 3.78 — Grey level co-occurrence matrices for the 0◦ (a), 45◦ (b), 90◦ (c) and 135◦ (d) directions. In vector notation these directions are m = (1, 0), m = (1, 1), m = (0, 1) and m = (−1, 1), respectively.

It should be noted that while a distance δ = 1 is commonly used for GLCM, other distances are possible. However, this does not change the number of For example, for δ = 3 (in 3D) the voxels at (0, 0, 3), (0, 3, 0), (3, 0, 0), (0, 3, 3), (0, 3, −3), (3, 0, 3), (3, 0, −3), (3, 3, 0), (3, −3, 0), (3, 3, 3), (3, 3, −3), (3, −3, 3) and (3, −3, −3) from the center voxel are considered.

Aggregating features To improve rotational invariance, GLCM feature values are computed by aggregating information from the diﬀerent underlying directional matrices 23. Five methods can be used to aggregate GLCMs and arrive at a single feature value. A schematic example is shown in Figure 3.3. A feature may be aggregated as follows:

1. Features are computed from each 2D directional matrix and averaged over 2D directions and slices (BTW3).

2. Features are computed from a single matrix after merging 2D directional matrices per slice, and then averaged over slices (SUJT).

3. Features are computed from a single matrix after merging 2D directional matrices per direction, and then averaged over directions (JJUI).

4. The feature is computed from a single matrix after merging all 2D directional matrices (ZW7Z).

5. Features are computed from each 3D directional matrix and averaged over the 3D directions (ITBB).

6. The feature is computed from a single matrix after merging all 3D directional matrices (IAZD).

In methods 2,3,4 and 6, matrices are merged by summing the co-occurrence counts in each matrix element (i, j) over the diﬀerent matrices. Probability distributions are subsequently calculated for the merged GLCM, which is then used to calculate GLCM features. Feature values may dependent strongly on the aggregation method. CHAPTER 3. IMAGE FEATURES 61

P j j j pi. 0 3 0 2 5 0.00 0.13 0.00 0.08 0.21 3 2 3 2 10 0.13 0.08 0.13 0.08 0.42 i i 0 3 2 0 5 0.00 0.13 0.08 0.00 0.21 2 2 0 0 4 0.08 0.08 0.00 0.00 0.17 P i 5 10 5 4 24 p.j 0.21 0.42 0.21 0.17 1.00

(a) Mm=(1,0) with margins (b) Pm=(1,0) with margins

k = |i − j| 0 1 2 3

pi−j 0.17 0.50 0.17 0.17

k = i + j 2 3 4 5 6 7 8

pi+j 0.00 0.25 0.08 0.42 0.25 0.00 0.00

(d) Cross-diagonal probability for Pm=(1,0) Table 3.79 — Grey level co-occurrence matrix for the 0◦ direction (a); its corresponding probability matrix Pm=(1,0) with marginal probabilities pi. and p.j(b); the diagonal probabilities pi−j (c); and the cross-diagonal probabilities pi+j (d). Discrepancies in panels b, c, and d are due to rounding errors caused by showing only two decimal places. Also, note that due to GLCM symmetry marginal probabilities pi. and p.j are the same in both row and column margins of panel b.

Distances and distance weighting The default neighbourhood includes all voxels within Chebyshev distance 1. The corresponding direction vectors are multiplied by the desired distance δ. From a technical point-of-view, direction vectors may also be determined differently, using any distance norm. In this case, direction vectors are the vectors to the voxels at δ, or between δ and δ − 1 for the Euclidean norm. Such usage is however rare and we caution against it due to potential reproducibility issues. GLCMs may be weighted for distance by multiplying M with a weighting factor w. By default w = 1, but w may also be an inverse distance function to weight each GLCM, e.g. w = kmk−1 or w = exp(−kmk2) 81, with kmk the length of direction vector m. Whether distance weighting yields different feature values depends on several factors. When aggregating the feature values, matrices have to be merged first, otherwise weighting has no effect. Also, it has no effect if the default neighbourhood is used and the Chebyshev norm is using for weighting. Nor does weighting have an effect if either Manhattan or Chebyshev norms are used both for constructing a non-default neighbourhood and for weighting. Weighting may furthermore have no effect for distance δ = 1, dependent on distance norms. Because of these exceptions, we recommend against using distance weighting for GLCM.

3.6.1 Joint maximum GYBY

Joint maximum 35 is the probability corresponding to the most common grey level co-occurrence in the GLCM:

Fcm.joint.max = max(pij) CHAPTER 3. IMAGE FEATURES 62

(a) 2D: by slice, without merging (b) 2D: by slice, with merging by slice

(e) 3D: as volume, without merging (f) 3D: as volume, with full merging

Figure 3.3 — Approaches to calculating grey level co-occurrence matrix-based features. M∆k are texture matrices calculated for direction ∆ in slice k (if applicable), and f∆k is the corresponding feature value. In (b-d) and (e) the matrices are merged prior to feature calculation.

data aggr. method value tol. consensus

dig. phantom 2D, averaged 0.519 — very strong dig. phantom 2D, slice-merged 0.512 — strong dig. phantom 2.5D, direction-merged 0.489 — strong dig. phantom 2.5D, merged 0.492 — strong dig. phantom 3D, averaged 0.503 — very strong dig. phantom 3D, merged 0.509 — very strong config. A 2D, averaged 0.109 0.001 strong config. A 2D, slice-merged 0.109 0.001 strong config. A 2.5D, direction-merged 0.0943 0.0008 strong config. A 2.5D, merged 0.0943 0.0008 strong config. B 2D, averaged 0.156 0.002 strong config. B 2D, slice-merged 0.156 0.002 strong config. B 2.5D, direction-merged 0.126 0.002 strong config. B 2.5D, merged 0.126 0.002 strong config. C 3D, averaged 0.111 0.002 strong config. C 3D, merged 0.111 0.002 very strong config. D 3D, averaged 0.232 0.007 strong config. D 3D, merged 0.232 0.007 strong config. E 3D, averaged 0.153 0.003 moderate CHAPTER 3. IMAGE FEATURES 63

conﬁg. E 3D, merged 0.153 0.003 strong

Table 3.80 — Reference values for the joint maximum feature.

3.6.2 Joint average 60VM

Joint average 74 is the grey level weighted sum of joint probabilities:

Ng Ng X X Fcm.joint.avg = i pij i=1 j=1

data aggr. method value tol. consensus

dig. phantom 2D, averaged 2.14 — very strong dig. phantom 2D, slice-merged 2.14 — strong dig. phantom 2.5D, direction-merged 2.2 — strong dig. phantom 2.5D, merged 2.2 — strong dig. phantom 3D, averaged 2.14 — very strong dig. phantom 3D, merged 2.15 — very strong config. A 2D, averaged 20.6 0.1 strong config. A 2D, slice-merged 20.6 0.1 strong config. A 2.5D, direction-merged 21.3 0.1 strong config. A 2.5D, merged 21.3 0.1 strong config. B 2D, averaged 18.7 0.3 strong config. B 2D, slice-merged 18.7 0.3 strong config. B 2.5D, direction-merged 19.2 0.3 strong config. B 2.5D, merged 19.2 0.3 strong config. C 3D, averaged 39 0.2 strong config. C 3D, merged 39 0.2 strong config. D 3D, averaged 18.9 0.5 strong config. D 3D, merged 18.9 0.5 strong config. E 3D, averaged 22.1 0.3 strong config. E 3D, merged 22.1 0.3 strong

Table 3.81 — Reference values for the joint average feature.

3.6.3 Joint variance UR99

The joint variance 74, which is also called sum of squares 36, is deﬁned as:

Ng Ng X X 2 Fcm.joint.var = (i − µ) pij i=1 j=1

Here µ is equal to the value of Fcm.joint.avg , which was deﬁned above.

data aggr. method value tol. consensus CHAPTER 3. IMAGE FEATURES 64

dig. phantom 2D, averaged 2.69 — very strong dig. phantom 2D, slice-merged 2.71 — strong dig. phantom 2.5D, direction-merged 3.22 — strong dig. phantom 2.5D, merged 3.24 — strong dig. phantom 3D, averaged 3.1 — very strong dig. phantom 3D, merged 3.13 — very strong config. A 2D, averaged 27 0.4 strong config. A 2D, slice-merged 27 0.4 strong config. A 2.5D, direction-merged 18.6 0.5 strong config. A 2.5D, merged 18.6 0.5 strong config. B 2D, averaged 21 0.3 strong config. B 2D, slice-merged 21 0.3 strong config. B 2.5D, direction-merged 14.2 0.1 strong config. B 2.5D, merged 14.2 0.1 strong config. C 3D, averaged 73.7 2 strong config. C 3D, merged 73.8 2 very strong config. D 3D, averaged 17.6 0.4 strong config. D 3D, merged 17.6 0.4 strong config. E 3D, averaged 24.4 0.9 moderate config. E 3D, merged 24.4 0.9 strong

Table 3.82 — Reference values for the joint variance feature.

3.6.4 Joint entropy TU9B

Joint entropy 36 is deﬁned as:

Ng Ng X X Fcm.joint.entr = − pij log2 pij i=1 j=1

data aggr. method value tol. consensus

dig. phantom 2D, averaged 2.05 — very strong dig. phantom 2D, slice-merged 2.24 — strong dig. phantom 2.5D, direction-merged 2.48 — strong dig. phantom 2.5D, merged 2.61 — strong dig. phantom 3D, averaged 2.4 — very strong dig. phantom 3D, merged 2.57 — very strong config. A 2D, averaged 5.82 0.04 strong config. A 2D, slice-merged 5.9 0.04 strong config. A 2.5D, direction-merged 5.78 0.04 strong config. A 2.5D, merged 5.79 0.04 strong config. B 2D, averaged 5.26 0.02 strong config. B 2D, slice-merged 5.45 0.01 strong config. B 2.5D, direction-merged 5.45 0.01 strong config. B 2.5D, merged 5.46 0.01 strong CHAPTER 3. IMAGE FEATURES 65

config. C 3D, averaged 6.39 0.06 strong config. C 3D, merged 6.42 0.06 very strong config. D 3D, averaged 4.95 0.03 strong config. D 3D, merged 4.96 0.03 strong config. E 3D, averaged 5.6 0.03 strong config. E 3D, merged 5.61 0.03 strong

Table 3.83 — Reference values for the joint entropy feature.

3.6.5 Diﬀerence average TF7R

The diﬀerence average 74 for the diagonal probabilities is deﬁned as:

Ng −1 X Fcm.diﬀ .avg = k pi−j,k k=0

By deﬁnition diﬀerence average is equivalent to the dissimilarity feature 81.

data aggr. method value tol. consensus

dig. phantom 2D, averaged 1.42 — very strong dig. phantom 2D, slice-merged 1.4 — strong dig. phantom 2.5D, direction-merged 1.46 — strong dig. phantom 2.5D, merged 1.44 — strong dig. phantom 3D, averaged 1.43 — very strong dig. phantom 3D, merged 1.38 — very strong config. A 2D, averaged 1.58 0.03 strong config. A 2D, slice-merged 1.57 0.03 strong config. A 2.5D, direction-merged 1.35 0.03 strong config. A 2.5D, merged 1.35 0.03 strong config. B 2D, averaged 1.81 0.01 strong config. B 2D, slice-merged 1.81 0.01 strong config. B 2.5D, direction-merged 1.47 0.01 strong config. B 2.5D, merged 1.47 0.01 strong config. C 3D, averaged 2.17 0.05 strong config. C 3D, merged 2.16 0.05 strong config. D 3D, averaged 1.29 0.01 strong config. D 3D, merged 1.29 0.01 strong config. E 3D, averaged 1.7 0.01 strong config. E 3D, merged 1.7 0.01 strong

Table 3.84 — Reference values for the diﬀerence average feature. CHAPTER 3. IMAGE FEATURES 66

3.6.6 Diﬀerence variance D3YU

The diﬀerence variance for the diagonal probabilities 36 is deﬁned as:

Ng −1 X 2 Fcm.diﬀ .var = (k − µ) pi−j,k k=0

Here µ is equal to the value of diﬀerence average.

data aggr. method value tol. consensus

dig. phantom 2D, averaged 2.9 — very strong dig. phantom 2D, slice-merged 3.06 — strong dig. phantom 2.5D, direction-merged 3.11 — strong dig. phantom 2.5D, merged 3.23 — strong dig. phantom 3D, averaged 3.06 — very strong dig. phantom 3D, merged 3.21 — very strong config. A 2D, averaged 4.94 0.19 strong config. A 2D, slice-merged 4.96 0.19 strong config. A 2.5D, direction-merged 4.12 0.2 strong config. A 2.5D, merged 4.14 0.2 strong config. B 2D, averaged 7.74 0.05 strong config. B 2D, slice-merged 7.76 0.05 strong config. B 2.5D, direction-merged 6.48 0.06 strong config. B 2.5D, merged 6.48 0.06 strong config. C 3D, averaged 14.4 0.5 strong config. C 3D, merged 14.4 0.5 strong config. D 3D, averaged 5.37 0.11 strong config. D 3D, merged 5.38 0.11 strong config. E 3D, averaged 8.22 0.06 strong config. E 3D, merged 8.23 0.06 strong

Table 3.85 — Reference values for the diﬀerence variance feature.

3.6.7 Diﬀerence entropy NTRS

The diﬀerence entropy for the diagonal probabilities 36 is deﬁned as:

Ng −1 X Fcm.diﬀ .entr = − pi−j,k log2 pi−j,k k=0

data aggr. method value tol. consensus

dig. phantom 2D, averaged 1.4 — very strong dig. phantom 2D, slice-merged 1.49 — strong dig. phantom 2.5D, direction-merged 1.61 — strong dig. phantom 2.5D, merged 1.67 — strong dig. phantom 3D, averaged 1.56 — very strong CHAPTER 3. IMAGE FEATURES 67

dig. phantom 3D, merged 1.64 — very strong config. A 2D, averaged 2.27 0.03 strong config. A 2D, slice-merged 2.28 0.03 strong config. A 2.5D, direction-merged 2.16 0.03 strong config. A 2.5D, merged 2.16 0.03 strong config. B 2D, averaged 2.35 0.01 strong config. B 2D, slice-merged 2.38 0.01 strong config. B 2.5D, direction-merged 2.24 0.01 moderate config. B 2.5D, merged 2.24 0.01 strong config. C 3D, averaged 2.64 0.03 strong config. C 3D, merged 2.64 0.03 very strong config. D 3D, averaged 2.13 0.01 strong config. D 3D, merged 2.14 0.01 strong config. E 3D, averaged 2.39 0.01 strong config. E 3D, merged 2.4 0.01 strong

Table 3.86 — Reference values for the diﬀerence entropy feature.

3.6.8 Sum average ZGXS

The sum average for the cross-diagonal probabilities 36 is deﬁned as:

2Ng X Fcm.sum.avg = k pi+j,k k=2

81 By deﬁnition, Fcm.sum.avg = 2Fcm.joint.avg .

data aggr. method value tol. consensus

dig. phantom 2D, averaged 4.28 — very strong dig. phantom 2D, slice-merged 4.29 — strong dig. phantom 2.5D, direction-merged 4.41 — strong dig. phantom 2.5D, merged 4.41 — strong dig. phantom 3D, averaged 4.29 — very strong dig. phantom 3D, merged 4.3 — very strong config. A 2D, averaged 41.3 0.1 strong config. A 2D, slice-merged 41.3 0.1 strong config. A 2.5D, direction-merged 42.7 0.1 strong config. A 2.5D, merged 42.7 0.1 strong config. B 2D, averaged 37.4 0.5 strong config. B 2D, slice-merged 37.4 0.5 strong config. B 2.5D, direction-merged 38.5 0.6 strong config. B 2.5D, merged 38.5 0.6 strong config. C 3D, averaged 78 0.3 strong config. C 3D, merged 78 0.3 strong config. D 3D, averaged 37.7 0.8 strong config. D 3D, merged 37.7 0.8 strong config. E 3D, averaged 44.3 0.4 strong CHAPTER 3. IMAGE FEATURES 68

conﬁg. E 3D, merged 44.3 0.4 strong

Table 3.87 — Reference values for the sum average feature.

3.6.9 Sum variance OEEB

The sum variance for the cross-diagonal probabilities 36 is deﬁned as:

2Ng X 2 Fcm.sum.var = (k − µ) pi+j,k k=2

Here µ is equal to the value of sum average. Sum variance is mathematically identical to the cluster tendency feature 81.

data aggr. method value tol. consensus

dig. phantom 2D, averaged 5.47 — very strong dig. phantom 2D, slice-merged 5.66 — strong dig. phantom 2.5D, direction-merged 7.48 — strong dig. phantom 2.5D, merged 7.65 — strong dig. phantom 3D, averaged 7.07 — very strong dig. phantom 3D, merged 7.41 — very strong config. A 2D, averaged 100 1 strong config. A 2D, slice-merged 100 1 strong config. A 2.5D, direction-merged 68.5 1.3 strong config. A 2.5D, merged 68.5 1.3 strong config. B 2D, averaged 72.1 1 strong config. B 2D, slice-merged 72.3 1 strong config. B 2.5D, direction-merged 48.1 0.4 strong config. B 2.5D, merged 48.1 0.4 strong config. C 3D, averaged 276 8 strong config. C 3D, merged 276 8 very strong config. D 3D, averaged 63.4 1.3 strong config. D 3D, merged 63.5 1.3 strong config. E 3D, averaged 86.6 3.3 moderate config. E 3D, merged 86.7 3.3 strong

Table 3.88 — Reference values for the sum variance feature.

3.6.10 Sum entropy P6QZ

The sum entropy for the cross-diagonal probabilities 36 is deﬁned as:

2Ng X Fcm.sum.entr = − pi+j,k log2 pi+j,k k=2

data aggr. method value tol. consensus CHAPTER 3. IMAGE FEATURES 69

dig. phantom 2D, averaged 1.6 — very strong dig. phantom 2D, slice-merged 1.79 — strong dig. phantom 2.5D, direction-merged 2.01 — strong dig. phantom 2.5D, merged 2.14 — strong dig. phantom 3D, averaged 1.92 — very strong dig. phantom 3D, merged 2.11 — very strong config. A 2D, averaged 4.19 0.03 strong config. A 2D, slice-merged 4.21 0.03 strong config. A 2.5D, direction-merged 4.17 0.03 strong config. A 2.5D, merged 4.18 0.03 strong config. B 2D, averaged 3.83 0.01 strong config. B 2D, slice-merged 3.89 0.01 strong config. B 2.5D, direction-merged 3.91 0.01 strong config. B 2.5D, merged 3.91 0.01 strong config. C 3D, averaged 4.56 0.04 strong config. C 3D, merged 4.56 0.04 very strong config. D 3D, averaged 3.68 0.02 strong config. D 3D, merged 3.68 0.02 strong config. E 3D, averaged 3.96 0.02 strong config. E 3D, merged 3.97 0.02 strong

Table 3.89 — Reference values for the sum entropy feature.

3.6.11 Angular second moment 8ZQL

36 The angular second moment , which represents the energy of P∆, is deﬁned as:

Ng Ng X X 2 Fcm.energy = pij i=1 j=1

This feature is also called energy 1,74 and uniformity 18.

data aggr. method value tol. consensus

dig. phantom 2D, averaged 0.368 — very strong dig. phantom 2D, slice-merged 0.352 — strong dig. phantom 2.5D, direction-merged 0.286 — strong dig. phantom 2.5D, merged 0.277 — strong dig. phantom 3D, averaged 0.303 — very strong dig. phantom 3D, merged 0.291 — very strong config. A 2D, averaged 0.045 0.0008 strong config. A 2D, slice-merged 0.0446 0.0008 strong config. A 2.5D, direction-merged 0.0429 0.0007 strong config. A 2.5D, merged 0.0427 0.0007 strong config. B 2D, averaged 0.0678 0.0006 strong config. B 2D, slice-merged 0.0669 0.0006 strong config. B 2.5D, direction-merged 0.0581 0.0006 strong config. B 2.5D, merged 0.058 0.0006 strong CHAPTER 3. IMAGE FEATURES 70

config. C 3D, averaged 0.045 0.001 strong config. C 3D, merged 0.0447 0.001 very strong config. D 3D, averaged 0.11 0.003 strong config. D 3D, merged 0.109 0.003 strong config. E 3D, averaged 0.0638 0.0009 strong config. E 3D, merged 0.0635 0.0009 strong

Table 3.90 — Reference values for the angular second moment feature.

3.6.12 Contrast ACUI

36 Contrast assesses grey level variations . Hence elements of M∆ that represent large grey level diﬀerences receive greater weight. Contrast is deﬁned as 18:

Ng Ng X X 2 Fcm.contrast = (i − j) pij i=1 j=1

Note that the original deﬁnition by Haralick et al. 36 is seemingly more complex, but rearranging and simplifying terms leads to the above formulation of contrast.

data aggr. method value tol. consensus

dig. phantom 2D, averaged 5.28 — very strong dig. phantom 2D, slice-merged 5.19 — strong dig. phantom 2.5D, direction-merged 5.39 — strong dig. phantom 2.5D, merged 5.29 — strong dig. phantom 3D, averaged 5.32 — very strong dig. phantom 3D, merged 5.12 — very strong config. A 2D, averaged 7.85 0.26 strong config. A 2D, slice-merged 7.82 0.26 strong config. A 2.5D, direction-merged 5.96 0.27 strong config. A 2.5D, merged 5.95 0.27 strong config. B 2D, averaged 11.9 0.1 strong config. B 2D, slice-merged 11.8 0.1 strong config. B 2.5D, direction-merged 8.66 0.09 strong config. B 2.5D, merged 8.65 0.09 strong config. C 3D, averaged 19.2 0.7 strong config. C 3D, merged 19.1 0.7 very strong config. D 3D, averaged 7.07 0.13 strong config. D 3D, merged 7.05 0.13 strong config. E 3D, averaged 11.1 0.1 strong config. E 3D, merged 11.1 0.1 strong

Table 3.91 — Reference values for the contrast feature. CHAPTER 3. IMAGE FEATURES 71

3.6.13 Dissimilarity 8S9J

Dissimilarity 18 is conceptually similar to the contrast feature, and is deﬁned as:

Ng Ng X X Fcm.dissimilarity = |i − j| pij i=1 j=1

By deﬁnition dissimilarity is equivalent to the diﬀerence average feature 81.

data aggr. method value tol. consensus

dig. phantom 2D, averaged 1.42 — very strong dig. phantom 2D, slice-merged 1.4 — strong dig. phantom 2.5D, direction-merged 1.46 — strong dig. phantom 2.5D, merged 1.44 — strong dig. phantom 3D, averaged 1.43 — very strong dig. phantom 3D, merged 1.38 — very strong config. A 2D, averaged 1.58 0.03 strong config. A 2D, slice-merged 1.57 0.03 strong config. A 2.5D, direction-merged 1.35 0.03 strong config. A 2.5D, merged 1.35 0.03 strong config. B 2D, averaged 1.81 0.01 strong config. B 2D, slice-merged 1.81 0.01 strong config. B 2.5D, direction-merged 1.47 0.01 strong config. B 2.5D, merged 1.47 0.01 strong config. C 3D, averaged 2.17 0.05 strong config. C 3D, merged 2.16 0.05 very strong config. D 3D, averaged 1.29 0.01 strong config. D 3D, merged 1.29 0.01 strong config. E 3D, averaged 1.7 0.01 strong config. E 3D, merged 1.7 0.01 strong

Table 3.92 — Reference values for the dissimilarity feature.

3.6.14 Inverse diﬀerence IB1Z

Inverse difference is a measure of homogeneity 18. Grey level co-occurrences with a large difference in levels are weighed less, thus lowering the total feature value. The feature score is maximal if all grey levels are the same. Inverse difference is defined as:

Ng Ng X X pij F = cm.inv.diﬀ 1 + |i − j| i=1 j=1

The equation above may also be expressed in terms of diagonal probabilities 81:

Ng −1 X pi−j,k F = cm.inv.diﬀ 1 + k k=0 CHAPTER 3. IMAGE FEATURES 72

data aggr. method value tol. consensus

dig. phantom 2D, averaged 0.678 — very strong dig. phantom 2D, slice-merged 0.683 — strong dig. phantom 2.5D, direction-merged 0.668 — strong dig. phantom 2.5D, merged 0.673 — strong dig. phantom 3D, averaged 0.677 — very strong dig. phantom 3D, merged 0.688 — very strong config. A 2D, averaged 0.581 0.003 strong config. A 2D, slice-merged 0.581 0.003 strong config. A 2.5D, direction-merged 0.605 0.003 strong config. A 2.5D, merged 0.605 0.003 strong config. B 2D, averaged 0.592 0.001 strong config. B 2D, slice-merged 0.593 0.001 strong config. B 2.5D, direction-merged 0.628 0.001 strong config. B 2.5D, merged 0.628 0.001 strong config. C 3D, averaged 0.582 0.004 strong config. C 3D, merged 0.583 0.004 very strong config. D 3D, averaged 0.682 0.003 strong config. D 3D, merged 0.682 0.003 strong config. E 3D, averaged 0.608 0.001 moderate config. E 3D, merged 0.608 0.001 strong

Table 3.93 — Reference values for the inverse diﬀerence feature.

3.6.15 Normalised inverse diﬀerence NDRX

Clausi 18 suggested normalising inverse difference to improve classification ability. The normalised feature is then defined as:

Ng Ng X X pij F = cm.inv.diﬀ .norm 1 + |i − j|/N i=1 j=1 g

2 2 Note that in Clausi’s definition, |i−j| /Ng is used instead of |i−j|/Ng, which is likely an oversight, as this exactly matches the definition of the normalised inverse difference moment feature. The equation may also be expressed in terms of diagonal probabilities 81:

Ng −1 X pi−j,k Fcm.inv.diﬀ .norm = 1 + k/Ng k=0

data aggr. method value tol. consensus

dig. phantom 2D, averaged 0.851 — very strong dig. phantom 2D, slice-merged 0.854 — strong dig. phantom 2.5D, direction-merged 0.847 — strong dig. phantom 2.5D, merged 0.85 — strong dig. phantom 3D, averaged 0.851 — very strong CHAPTER 3. IMAGE FEATURES 73

dig. phantom 3D, merged 0.856 — very strong config. A 2D, averaged 0.961 0.001 strong config. A 2D, slice-merged 0.961 0.001 strong config. A 2.5D, direction-merged 0.966 0.001 strong config. A 2.5D, merged 0.966 0.001 strong config. B 2D, averaged 0.952 0.001 strong config. B 2D, slice-merged 0.952 0.001 strong config. B 2.5D, direction-merged 0.96 0.001 strong config. B 2.5D, merged 0.96 0.001 strong config. C 3D, averaged 0.966 0.001 strong config. C 3D, merged 0.966 0.001 very strong config. D 3D, averaged 0.965 0.001 strong config. D 3D, merged 0.965 0.001 strong config. E 3D, averaged 0.955 0.001 strong config. E 3D, merged 0.955 0.001 strong

Table 3.94 — Reference values for the normalised inverse diﬀerence feature.

3.6.16 Inverse diﬀerence moment WF0Z

Inverse diﬀerence moment 36 is similar in concept to the inverse diﬀerence feature, but with lower weights for elements that are further from the diagonal:

Ng Ng X X pij Fcm.inv.diﬀ .mom = 2 i=1 j=1 1 + (i − j)

The equation above may also be expressed in terms of diagonal probabilities 81:

Ng −1 X pi−j,k F = cm.inv.diﬀ .mom 1 + k2 k=0

This feature is also called homogeneity 74.

data aggr. method value tol. consensus

dig. phantom 2D, averaged 0.619 — very strong dig. phantom 2D, slice-merged 0.625 — strong dig. phantom 2.5D, direction-merged 0.606 — strong dig. phantom 2.5D, merged 0.613 — strong dig. phantom 3D, averaged 0.618 — very strong dig. phantom 3D, merged 0.631 — very strong config. A 2D, averaged 0.544 0.003 strong config. A 2D, slice-merged 0.544 0.003 strong config. A 2.5D, direction-merged 0.573 0.003 strong config. A 2.5D, merged 0.573 0.003 strong config. B 2D, averaged 0.557 0.001 strong config. B 2D, slice-merged 0.558 0.001 strong CHAPTER 3. IMAGE FEATURES 74

config. B 2.5D, direction-merged 0.6 0.001 strong config. B 2.5D, merged 0.6 0.001 strong config. C 3D, averaged 0.547 0.004 strong config. C 3D, merged 0.548 0.004 very strong config. D 3D, averaged 0.656 0.003 strong config. D 3D, merged 0.657 0.003 strong config. E 3D, averaged 0.576 0.001 strong config. E 3D, merged 0.577 0.001 strong

Table 3.95 — Reference values for the inverse diﬀerence moment feature.

3.6.17 Normalised inverse diﬀerence moment 1QCO

Clausi 18 suggested normalising inverse difference moment to improve classification performance. This leads to the following definition:

Ng Ng X X pij Fcm.inv.diﬀ .mom.norm = 2 2 i=1 j=1 1 + (i − j) /Ng

The equation above may also be expressed in terms of diagonal probabilities 81:

Ng −1 X pi−j,k Fcm.inv.diﬀ .mom.norm = 2 k=0 1 + (k/Ng)

data aggr. method value tol. consensus

dig. phantom 2D, averaged 0.899 — very strong dig. phantom 2D, slice-merged 0.901 — strong dig. phantom 2.5D, direction-merged 0.897 — strong dig. phantom 2.5D, merged 0.899 — strong dig. phantom 3D, averaged 0.898 — very strong dig. phantom 3D, merged 0.902 — very strong config. A 2D, averaged 0.994 0.001 strong config. A 2D, slice-merged 0.994 0.001 strong config. A 2.5D, direction-merged 0.996 0.001 strong config. A 2.5D, merged 0.996 0.001 strong config. B 2D, averaged 0.99 0.001 strong config. B 2D, slice-merged 0.99 0.001 strong config. B 2.5D, direction-merged 0.992 0.001 strong config. B 2.5D, merged 0.992 0.001 strong config. C 3D, averaged 0.994 0.001 strong config. C 3D, merged 0.994 0.001 very strong config. D 3D, averaged 0.994 0.001 strong config. D 3D, merged 0.994 0.001 strong config. E 3D, averaged 0.99 0.001 strong config. E 3D, merged 0.99 0.001 strong CHAPTER 3. IMAGE FEATURES 75

Table 3.96 — Reference values for the normalised inverse diﬀerence moment feature.

3.6.18 Inverse variance E8JP

The inverse variance 1 feature is deﬁned as:

Ng Ng X X pij Fcm.inv.var = 2 2 i=1 j>i (i − j)

The equation above may also be expressed in terms of diagonal probabilities. Note that in this case, summation starts at k = 1 instead of k = 0 81:

Ng −1 X pi−j,k F = cm.inv.var k2 k=1

data aggr. method value tol. consensus

dig. phantom 2D, averaged 0.0567 — very strong dig. phantom 2D, slice-merged 0.0553 — strong dig. phantom 2.5D, direction-merged 0.0597 — strong dig. phantom 2.5D, merged 0.0582 — strong dig. phantom 3D, averaged 0.0604 — very strong dig. phantom 3D, merged 0.0574 — very strong config. A 2D, averaged 0.441 0.001 strong config. A 2D, slice-merged 0.441 0.001 strong config. A 2.5D, direction-merged 0.461 0.002 strong config. A 2.5D, merged 0.461 0.002 strong config. B 2D, averaged 0.401 0.002 strong config. B 2D, slice-merged 0.401 0.002 strong config. B 2.5D, direction-merged 0.424 0.003 strong config. B 2.5D, merged 0.424 0.003 strong config. C 3D, averaged 0.39 0.003 strong config. C 3D, merged 0.39 0.003 very strong config. D 3D, averaged 0.341 0.005 strong config. D 3D, merged 0.34 0.005 strong config. E 3D, averaged 0.41 0.004 strong config. E 3D, merged 0.41 0.004 strong

Table 3.97 — Reference values for the inverse variance feature.

3.6.19 Correlation NI2N

Correlation 36 is deﬁned as:

 Ng Ng  1 X X F = −µ µ + i j p cm.corr σ σ  i. .j ij i. .j i=1 j=1 CHAPTER 3. IMAGE FEATURES 76

1/2 PNg PNg 2 µi. = i=1 i pi. and σi. = i=1(i − µi.) pi. are the mean and standard deviation of row marginal probability pi., respectively. Likewise, µ.j and σ.j are the mean and standard deviation of the column marginal probability p.j, respectively. The calculation of correlation can be simpliﬁed since P∆ is symmetrical:

 Ng Ng  1 X X F = −µ2 + i j p cm.corr σ2  i. ij i. i=1 j=1

An equivalent formulation of correlation is:

Ng Ng 1 X X F = (i − µ )(j − µ ) p cm.corr σ σ i. .j ij i. .j i=1 j=1

Again, simplifying due to matrix symmetry yields:

Ng Ng 1 X X F = (i − µ )(j − µ ) p cm.corr σ2 i. i. ij i. i=1 j=1

data aggr. method value tol. consensus

dig. phantom 2D, averaged −0.0121 — very strong dig. phantom 2D, slice-merged 0.0173 — strong dig. phantom 2.5D, direction-merged 0.178 — strong dig. phantom 2.5D, merged 0.182 — strong dig. phantom 3D, averaged 0.157 — very strong dig. phantom 3D, merged 0.183 — very strong config. A 2D, averaged 0.778 0.002 strong config. A 2D, slice-merged 0.78 0.002 strong config. A 2.5D, direction-merged 0.839 0.003 strong config. A 2.5D, merged 0.84 0.003 strong config. B 2D, averaged 0.577 0.002 strong config. B 2D, slice-merged 0.58 0.002 strong config. B 2.5D, direction-merged 0.693 0.003 strong config. B 2.5D, merged 0.695 0.003 strong config. C 3D, averaged 0.869 0.001 strong config. C 3D, merged 0.871 0.001 strong config. D 3D, averaged 0.798 0.005 strong config. D 3D, merged 0.8 0.005 strong config. E 3D, averaged 0.771 0.006 moderate config. E 3D, merged 0.773 0.006 strong

Table 3.98 — Reference values for the correlation feature. CHAPTER 3. IMAGE FEATURES 77

3.6.20 Autocorrelation QWB0

Soh and Tsatsoulis 63 deﬁned autocorrelation as:

Ng Ng X X Fcm.auto.corr = i j pij i=1 j=1

data aggr. method value tol. consensus

dig. phantom 2D, averaged 5.09 — very strong dig. phantom 2D, slice-merged 5.14 — strong dig. phantom 2.5D, direction-merged 5.4 — strong dig. phantom 2.5D, merged 5.45 — strong dig. phantom 3D, averaged 5.06 — very strong dig. phantom 3D, merged 5.19 — very strong config. A 2D, averaged 455 2 strong config. A 2D, slice-merged 455 2 strong config. A 2.5D, direction-merged 471 2 strong config. A 2.5D, merged 471 2 strong config. B 2D, averaged 369 11 strong config. B 2D, slice-merged 369 11 strong config. B 2.5D, direction-merged 380 11 strong config. B 2.5D, merged 380 11 strong config. C 3D, averaged 1.58 × 103 10 strong config. C 3D, merged 1.58 × 103 10 strong config. D 3D, averaged 370 16 strong config. D 3D, merged 370 16 strong config. E 3D, averaged 509 8 strong config. E 3D, merged 509 8 strong

Table 3.99 — Reference values for the autocorrelation feature.

3.6.21 Cluster tendency DG8W

Cluster tendency 1 is deﬁned as:

Ng Ng X X 2 Fcm.clust.tend = (i + j − µi. − µ.j) pij i=1 j=1

PNg PNg Here µi. = i=1 i pi. and µ.j = j=1 j p.j. Because of the symmetric nature of P∆, the feature can also be formulated as:

Ng Ng X X 2 Fcm.clust.tend = (i + j − 2µi.) pij i=1 j=1

Cluster tendency is mathematically equal to the sum variance feature 81. CHAPTER 3. IMAGE FEATURES 78

data aggr. method value tol. consensus

Table 3.100 — Reference values for the cluster tendency feature.

3.6.22 Cluster shade 7NFM

Cluster shade 74 is deﬁned as:

Ng Ng X X 3 Fcm.clust.shade = (i + j − µi. − µ.j) pij i=1 j=1

PNg PNg As with cluster tendency, µi. = i=1 i pi. and µ.j = j=1 j p.j. Because of the symmetric nature of P∆, the feature can also be formulated as:

Ng Ng X X 3 Fcm.clust.shade = (i + j − 2µi.) pij i=1 j=1

data aggr. method value tol. consensus

dig. phantom 2D, averaged 7 — very strong dig. phantom 2D, slice-merged 6.98 — strong dig. phantom 2.5D, direction-merged 16.6 — strong dig. phantom 2.5D, merged 16.4 — strong dig. phantom 3D, averaged 16.6 — very strong dig. phantom 3D, merged 17.4 — very strong conﬁg. A 2D, averaged −1.04 × 103 20 strong CHAPTER 3. IMAGE FEATURES 79

config. A 2D, slice-merged −1.05 × 103 20 strong config. A 2.5D, direction-merged −1.49 × 103 30 strong config. A 2.5D, merged −1.49 × 103 30 strong config. B 2D, averaged −668 17 strong config. B 2D, slice-merged −673 17 strong config. B 2.5D, direction-merged −905 19 strong config. B 2.5D, merged −906 19 strong config. C 3D, averaged −1.06 × 104 300 strong config. C 3D, merged −1.06 × 104 300 very strong config. D 3D, averaged −1.27 × 103 40 strong config. D 3D, merged −1.28 × 103 40 strong config. E 3D, averaged −2.07 × 103 70 moderate config. E 3D, merged −2.08 × 103 70 strong

Table 3.101 — Reference values for the cluster shade feature.

3.6.23 Cluster prominence AE86

Cluster prominence 74 is deﬁned as:

Ng Ng X X 4 Fcm.clust.prom = (i + j − µi. − µ.j) pij i=1 j=1

PNg PNg As before, µi. = i=1 i pi. and µ.j = j=1 j p.j. Because of the symmetric nature of P∆, the feature can also be formulated as:

Ng Ng X X 4 Fcm.clust.prom = (i + j − 2µi.) pij i=1 j=1

data aggr. method value tol. consensus

dig. phantom 2D, averaged 79.1 — very strong dig. phantom 2D, slice-merged 80.4 — strong dig. phantom 2.5D, direction-merged 147 — strong dig. phantom 2.5D, merged 142 — strong dig. phantom 3D, averaged 145 — very strong dig. phantom 3D, merged 147 — very strong config. A 2D, averaged 5.27 × 104 500 strong config. A 2D, slice-merged 5.28 × 104 500 strong config. A 2.5D, direction-merged 4.76 × 104 700 strong config. A 2.5D, merged 4.77 × 104 700 strong config. B 2D, averaged 2.94 × 104 1.4 × 103 strong config. B 2D, slice-merged 2.95 × 104 1.4 × 103 strong config. B 2.5D, direction-merged 2.52 × 104 1 × 103 strong config. B 2.5D, merged 2.53 × 104 1 × 103 strong config. C 3D, averaged 5.69 × 105 1.1 × 104 strong config. C 3D, merged 5.7 × 105 1.1 × 104 very strong CHAPTER 3. IMAGE FEATURES 80

config. D 3D, averaged 3.57 × 104 1.4 × 103 strong config. D 3D, merged 3.57 × 104 1.5 × 103 strong config. E 3D, averaged 6.89 × 104 2.1 × 103 moderate config. E 3D, merged 6.9 × 104 2.1 × 103 strong

Table 3.102 — Reference values for the cluster prominence feature.

3.6.24 Information correlation 1 R8DG

36 Information theoretic correlation is estimated using two different measures . For symmetric P∆ the first measure is defined as: HXY − HXY F = 1 cm.info.corr.1 HX

PNg PNg PNg HXY = − i=1 j=1 pij log2 pij is the entropy for the joint probability. HX = − i=1 pi. log2 pi. is the entropy for the row marginal probability, which due to symmetry is equal to the entropy of the column marginal probability. HXY 1 is a type of entropy that is deﬁned as:

Ng Ng X X HXY 1 = − pij log2 (pi.p.j) i=1 j=1

data aggr. method value tol. consensus

dig. phantom 2D, averaged −0.155 — very strong dig. phantom 2D, slice-merged −0.0341 — strong dig. phantom 2.5D, direction-merged −0.124 — strong dig. phantom 2.5D, merged −0.0334 — strong dig. phantom 3D, averaged −0.157 — very strong dig. phantom 3D, merged −0.0288 — very strong config. A 2D, averaged −0.236 0.001 strong config. A 2D, slice-merged −0.214 0.001 strong config. A 2.5D, direction-merged −0.231 0.001 strong config. A 2.5D, merged −0.228 0.001 strong config. B 2D, averaged −0.239 0.001 strong config. B 2D, slice-merged −0.181 0.001 strong config. B 2.5D, direction-merged −0.188 0.001 strong config. B 2.5D, merged −0.185 0.001 strong config. C 3D, averaged −0.236 0.001 strong config. C 3D, merged −0.228 0.001 strong config. D 3D, averaged −0.231 0.003 strong config. D 3D, merged −0.225 0.003 strong config. E 3D, averaged −0.181 0.003 moderate config. E 3D, merged −0.175 0.003 strong

Table 3.103 — Reference values for the information correlation 1 feature. CHAPTER 3. IMAGE FEATURES 81

3.6.25 Information correlation 2 JN9H

The second measure of information theoretic correlation 36 is estimated as follows for symmetric

P∆: p Fcm.info.corr.2 = 1 − exp (−2 (HXY 2 − HXY ))

PNg PNg As earlier, HXY = − i=1 j=1 pij log2 pij. HXY 2 is a type of entropy deﬁned as:

Ng Ng X X HXY 2 = − pi.p.j log2 (pi.p.j) i=1 j=1

data aggr. method value tol. consensus

dig. phantom 2D, averaged 0.487 — strong dig. phantom 2D, slice-merged 0.263 — strong dig. phantom 2.5D, direction-merged 0.487 — strong dig. phantom 2.5D, merged 0.291 — strong dig. phantom 3D, averaged 0.52 — very strong dig. phantom 3D, merged 0.269 — very strong config. A 2D, averaged 0.863 0.003 strong config. A 2D, slice-merged 0.851 0.002 strong config. A 2.5D, direction-merged 0.879 0.001 strong config. A 2.5D, merged 0.88 0.001 strong config. B 2D, averaged 0.837 0.001 strong config. B 2D, slice-merged 0.792 0.001 strong config. B 2.5D, direction-merged 0.821 0.001 strong config. B 2.5D, merged 0.819 0.001 strong config. C 3D, averaged 0.9 0.001 strong config. C 3D, merged 0.899 0.001 strong config. D 3D, averaged 0.845 0.003 strong config. D 3D, merged 0.846 0.003 strong config. E 3D, averaged 0.813 0.004 moderate config. E 3D, merged 0.813 0.004 strong

Table 3.104 — Reference values for the information correlation 2 feature. CHAPTER 3. IMAGE FEATURES 82

3.7 Grey level run length based features TP0I

The grey level run length matrix (GLRLM) was introduced by Galloway 30 to define various texture features. Like the grey level co-occurrence matrix, GLRLM also assesses the distribution of discretised grey levels in an image or in a stack of images. However, whereas GLCM assesses co-occurrence of grey levels within neighbouring pixels or voxels, GLRLM assesses run lengths. A run length is defined as the length of a consecutive sequence of pixels or voxels with the same grey level along direction m, which was previously defined in Section 3.6. The GLRLM then contains the occurrences of runs with length j for a discretised grey level i. A complete example for GLRLM construction from a 2D image is shown in Table 3.105. Let

Mm be the Ng × Nr grey level run length matrix, where Ng is the number of discretised grey levels present in the ROI intensity mask and Nr the maximal possible run length along direction m. Matrix element rij of the GLRLM is the occurrence of grev level i with run length j. Then, PNg PNr let Nv be the total number of voxels in the ROI intensity mask, and Ns = i=1 j=1 rij the sum over all elements in Mm. Marginal sums are also deﬁned. Let ri. be the marginal sum of the runs PNr over run lengths j for grey value i, that is ri. = j=1 rij. Similarly, the marginal sum of the runs PNg over the grey values i for run length j is r.j = i=1 rij.

Aggregating features To improve rotational invariance, GLRLM feature values are computed by aggregating information from the diﬀerent underlying directional matrices 23. Five methods can be used to aggregate GLRLMs and arrive at a single feature value. A schematic example was previously shown in Figure 3.3. A feature may be aggregated as follows:

1. Features are computed from each 2D directional matrix and averaged over 2D directions and slices (BTW3).

2. Features are computed from a single matrix after merging 2D directional matrices per slice, and then averaged over slices (SUJT).

3. Features are computed from a single matrix after merging 2D directional matrices per direction, and then averaged over directions (JJUI).

4. The feature is computed from a single matrix after merging all 2D directional matrices (ZW7Z).

5. Features are computed from each 3D directional matrix and averaged over the 3D directions (ITBB).

6. The feature is computed from a single matrix after merging all 3D directional matrices (IAZD).

In methods 2,3,4 and 6 matrices are merged by summing the run counts of each matrix element

(i, j) over the diﬀerent matrices. Note that when matrices are merged, Nv should likewise be summed to retain consistency. Feature values may dependent strongly on the aggregation method.

Distance weighting GLRLMs may be weighted for distance by multiplying the run lengths with a weighting factor w. By default w = 1, but w may also be an inverse distance function, e.g. w = kmk−1 or w = exp(−kmk2) 81, with kmk the length of direction vector m. Whether distance weighting yields different feature values depends on several factors. When aggregating the feature values, matrices have to be merged first, otherwise weighting has no effect. It also has no effect if the Chebyshev norm is used for weighting. Distance weighting is non-standard use, and we caution against it due to potential reproducibility issues. CHAPTER 3. IMAGE FEATURES 83

1 2 2 3 1 2 3 3 4 2 4 1 4 1 2 3

(a) Grey levels

Run length j Run length j 1 2 3 4 1 2 3 4 1 4 0 0 0 1 4 0 0 0 2 3 1 0 0 2 3 1 0 0 i i 3 2 1 0 0 3 2 1 0 0 4 3 0 0 0 4 3 0 0 0

(b) Mm=→ (c) Mm=%

Run length j Run length j 1 2 3 4 1 2 3 4 1 2 1 0 0 1 4 0 0 0 2 2 0 1 0 2 3 1 0 0 i i 3 2 1 0 0 3 4 0 0 0 4 1 1 0 0 4 3 0 0 0

(d) Mm=↑ (e) Mm=- Table 3.105 — Grey level run length matrices for the 0◦ (a), 45◦ (b), 90◦ (c) and 135◦ (d) directions. In vector notation these directions are m = (1, 0), m = (1, 1), m = (0, 1) and m = (−1, 1), respectively.

3.7.1 Short runs emphasis 22OV

This feature emphasises short run lengths 30. It is deﬁned as:

Nr 1 X r.j F = rlm.sre N j2 s j=1

data aggr. method value tol. consensus

dig. phantom 2D, averaged 0.641 — very strong dig. phantom 2D, slice-merged 0.661 — strong dig. phantom 2.5D, direction-merged 0.665 — strong dig. phantom 2.5D, merged 0.68 — strong dig. phantom 3D, averaged 0.705 — very strong dig. phantom 3D, merged 0.729 — very strong config. A 2D, averaged 0.785 0.003 strong config. A 2D, slice-merged 0.786 0.003 strong config. A 2.5D, direction-merged 0.768 0.003 strong config. A 2.5D, merged 0.769 0.003 strong CHAPTER 3. IMAGE FEATURES 84

config. B 2D, averaged 0.781 0.001 strong config. B 2D, slice-merged 0.782 0.001 strong config. B 2.5D, direction-merged 0.759 0.001 strong config. B 2.5D, merged 0.759 0.001 strong config. C 3D, averaged 0.786 0.003 strong config. C 3D, merged 0.787 0.003 strong config. D 3D, averaged 0.734 0.001 strong config. D 3D, merged 0.736 0.001 strong config. E 3D, averaged 0.776 0.001 moderate config. E 3D, merged 0.777 0.001 strong

Table 3.106 — Reference values for the short runs emphasis feature.

3.7.2 Long runs emphasis W4KF

This feature emphasises long run lengths 30. It is deﬁned as:

N 1 Xr F = j2r rlm.lre N .j s j=1

data aggr. method value tol. consensus

dig. phantom 2D, averaged 3.78 — very strong dig. phantom 2D, slice-merged 3.51 — strong dig. phantom 2.5D, direction-merged 3.46 — strong dig. phantom 2.5D, merged 3.27 — strong dig. phantom 3D, averaged 3.06 — very strong dig. phantom 3D, merged 2.76 — very strong config. A 2D, averaged 2.91 0.03 strong config. A 2D, slice-merged 2.89 0.03 strong config. A 2.5D, direction-merged 3.09 0.03 strong config. A 2.5D, merged 3.08 0.03 strong config. B 2D, averaged 3.52 0.04 strong config. B 2D, slice-merged 3.5 0.04 strong config. B 2.5D, direction-merged 3.82 0.05 strong config. B 2.5D, merged 3.81 0.05 strong config. C 3D, averaged 3.31 0.04 strong config. C 3D, merged 3.28 0.04 strong config. D 3D, averaged 6.66 0.18 strong config. D 3D, merged 6.56 0.18 strong config. E 3D, averaged 3.55 0.07 strong config. E 3D, merged 3.52 0.07 strong

Table 3.107 — Reference values for the long runs emphasis feature. CHAPTER 3. IMAGE FEATURES 85

3.7.3 Low grey level run emphasis V3SW

This feature is a grey level analogue to short runs emphasis 15. Instead of short run lengths, low grey levels are emphasised. The feature is deﬁned as:

Ng 1 X ri. F = rlm.lgre N i2 s i=1

data aggr. method value tol. consensus

dig. phantom 2D, averaged 0.604 — very strong dig. phantom 2D, slice-merged 0.609 — strong dig. phantom 2.5D, direction-merged 0.58 — strong dig. phantom 2.5D, merged 0.585 — strong dig. phantom 3D, averaged 0.603 — very strong dig. phantom 3D, merged 0.607 — very strong config. A 2D, averaged 0.0264 0.0003 strong config. A 2D, slice-merged 0.0264 0.0003 strong config. A 2.5D, direction-merged 0.0148 0.0004 strong config. A 2.5D, merged 0.0147 0.0004 strong config. B 2D, averaged 0.0331 0.0006 strong config. B 2D, slice-merged 0.033 0.0006 strong config. B 2.5D, direction-merged 0.0194 0.0006 strong config. B 2.5D, merged 0.0194 0.0006 strong config. C 3D, averaged 0.00155 5 × 10−5 strong config. C 3D, merged 0.00155 5 × 10−5 strong config. D 3D, averaged 0.0257 0.0012 strong config. D 3D, merged 0.0257 0.0012 strong config. E 3D, averaged 0.0204 0.0008 moderate config. E 3D, merged 0.0204 0.0008 strong

Table 3.108 — Reference values for the low grey level run emphasis feature.

3.7.4 High grey level run emphasis G3QZ

The high grey level run emphasis feature is a grey level analogue to long runs emphasis 15. The feature emphasises high grey levels, and is deﬁned as:

Ng 1 X F = i2r rlm.hgre N i. s i=1

data aggr. method value tol. consensus

dig. phantom 2D, averaged 9.82 — very strong dig. phantom 2D, slice-merged 9.74 — strong dig. phantom 2.5D, direction-merged 10.3 — strong dig. phantom 2.5D, merged 10.2 — strong CHAPTER 3. IMAGE FEATURES 86

dig. phantom 3D, averaged 9.7 — very strong dig. phantom 3D, merged 9.64 — very strong config. A 2D, averaged 428 3 strong config. A 2D, slice-merged 428 3 strong config. A 2.5D, direction-merged 449 3 strong config. A 2.5D, merged 449 3 strong config. B 2D, averaged 342 11 strong config. B 2D, slice-merged 342 11 strong config. B 2.5D, direction-merged 356 11 strong config. B 2.5D, merged 356 11 strong config. C 3D, averaged 1.47 × 103 10 strong config. C 3D, merged 1.47 × 103 10 strong config. D 3D, averaged 326 17 strong config. D 3D, merged 326 17 strong config. E 3D, averaged 471 9 strong config. E 3D, merged 471 9 strong

Table 3.109 — Reference values for the high grey level run emphasis feature.

3.7.5 Short run low grey level emphasis HTZT

This feature emphasises runs in the upper left quadrant of the GLRLM, where short run lengths and low grey levels are located 22. It is deﬁned as:

Ng Nr 1 X X rij F = rlm.srlge N i2j2 s i=1 j=1

data aggr. method value tol. consensus

dig. phantom 2D, averaged 0.294 — very strong dig. phantom 2D, slice-merged 0.311 — strong dig. phantom 2.5D, direction-merged 0.296 — strong dig. phantom 2.5D, merged 0.312 — strong dig. phantom 3D, averaged 0.352 — very strong dig. phantom 3D, merged 0.372 — very strong config. A 2D, averaged 0.0243 0.0003 strong config. A 2D, slice-merged 0.0243 0.0003 strong config. A 2.5D, direction-merged 0.0135 0.0004 strong config. A 2.5D, merged 0.0135 0.0004 strong config. B 2D, averaged 0.0314 0.0006 strong config. B 2D, slice-merged 0.0313 0.0006 strong config. B 2.5D, direction-merged 0.0181 0.0006 strong config. B 2.5D, merged 0.0181 0.0006 strong config. C 3D, averaged 0.00136 5 × 10−5 strong config. C 3D, merged 0.00136 5 × 10−5 strong config. D 3D, averaged 0.0232 0.001 strong CHAPTER 3. IMAGE FEATURES 87

config. D 3D, merged 0.0232 0.001 strong config. E 3D, averaged 0.0187 0.0007 moderate config. E 3D, merged 0.0186 0.0007 strong

Table 3.110 — Reference values for the short run low grey level emphasis feature.

3.7.6 Short run high grey level emphasis GD3A

This feature emphasises runs in the lower left quadrant of the GLRLM, where short run lengths and high grey levels are located 22. The feature is deﬁned as:

Ng Nr 2 1 X X i rij F = rlm.srhge N j2 s i=1 j=1

data aggr. method value tol. consensus

dig. phantom 2D, averaged 8.57 — very strong dig. phantom 2D, slice-merged 8.67 — strong dig. phantom 2.5D, direction-merged 9.03 — strong dig. phantom 2.5D, merged 9.05 — strong dig. phantom 3D, averaged 8.54 — very strong dig. phantom 3D, merged 8.67 — very strong config. A 2D, averaged 320 1 strong config. A 2D, slice-merged 320 1 strong config. A 2.5D, direction-merged 332 1 strong config. A 2.5D, merged 333 1 strong config. B 2D, averaged 251 8 strong config. B 2D, slice-merged 252 8 strong config. B 2.5D, direction-merged 257 9 strong config. B 2.5D, merged 258 9 strong config. C 3D, averaged 1.1 × 103 10 strong config. C 3D, merged 1.1 × 103 10 strong config. D 3D, averaged 219 13 strong config. D 3D, merged 219 13 strong config. E 3D, averaged 346 7 strong config. E 3D, merged 347 7 strong

Table 3.111 — Reference values for the short run high grey level emphasis feature.

3.7.7 Long run low grey level emphasis IVPO

This feature emphasises runs in the upper right quadrant of the GLRLM, where long run lengths and low grey levels are located 22. The feature is deﬁned as:

Ng Nr 2 1 X X j rij F = rlm.lrlge N i2 s i=1 j=1 CHAPTER 3. IMAGE FEATURES 88

data aggr. method value tol. consensus

dig. phantom 2D, averaged 3.14 — very strong dig. phantom 2D, slice-merged 2.92 — strong dig. phantom 2.5D, direction-merged 2.79 — strong dig. phantom 2.5D, merged 2.63 — strong dig. phantom 3D, averaged 2.39 — very strong dig. phantom 3D, merged 2.16 — very strong config. A 2D, averaged 0.0386 0.0003 strong config. A 2D, slice-merged 0.0385 0.0003 strong config. A 2.5D, direction-merged 0.0229 0.0004 strong config. A 2.5D, merged 0.0228 0.0004 strong config. B 2D, averaged 0.0443 0.0008 strong config. B 2D, slice-merged 0.0442 0.0008 strong config. B 2.5D, direction-merged 0.0293 0.0009 strong config. B 2.5D, merged 0.0292 0.0009 strong config. C 3D, averaged 0.00317 4 × 10−5 strong config. C 3D, merged 0.00314 4 × 10−5 strong config. D 3D, averaged 0.0484 0.0031 strong config. D 3D, merged 0.0478 0.0031 strong config. E 3D, averaged 0.0313 0.0016 moderate config. E 3D, merged 0.0311 0.0016 strong

Table 3.112 — Reference values for the long run low grey level emphasis feature.

3.7.8 Long run high grey level emphasis 3KUM

This feature emphasises runs in the lower right quadrant of the GLRLM, where long run lengths and high grey levels are located 22. The feature is deﬁned as:

Ng N 1 X Xr F = i2j2r rlm.lrhge N ij s i=1 j=1

data aggr. method value tol. consensus

dig. phantom 2D, averaged 17.4 — very strong dig. phantom 2D, slice-merged 16.1 — strong dig. phantom 2.5D, direction-merged 17.9 — strong dig. phantom 2.5D, merged 17 — strong dig. phantom 3D, averaged 17.6 — very strong dig. phantom 3D, merged 15.6 — very strong config. A 2D, averaged 1.41 × 103 20 strong config. A 2D, slice-merged 1.4 × 103 20 strong config. A 2.5D, direction-merged 1.5 × 103 20 strong config. A 2.5D, merged 1.5 × 103 20 strong config. B 2D, averaged 1.39 × 103 30 strong CHAPTER 3. IMAGE FEATURES 89

config. B 2D, slice-merged 1.38 × 103 30 strong config. B 2.5D, direction-merged 1.5 × 103 30 strong config. B 2.5D, merged 1.5 × 103 30 strong config. C 3D, averaged 5.59 × 103 80 strong config. C 3D, merged 5.53 × 103 80 strong config. D 3D, averaged 2.67 × 103 30 strong config. D 3D, merged 2.63 × 103 30 strong config. E 3D, averaged 1.9 × 103 20 moderate config. E 3D, merged 1.89 × 103 20 strong

Table 3.113 — Reference values for the long run high grey level emphasis feature.

3.7.9 Grey level non-uniformity R5YN

This feature assesses the distribution of runs over the grey values 30. The feature value is low when runs are equally distributed along grey levels. The feature is deﬁned as:

Ng 1 X F = r2 rlm.glnu N i. s i=1

data aggr. method value tol. consensus

dig. phantom 2D, averaged 5.2 — very strong dig. phantom 2D, slice-merged 20.5 — strong dig. phantom 2.5D, direction-merged 19.5 — strong dig. phantom 2.5D, merged 77.1 — strong dig. phantom 3D, averaged 21.8 — very strong dig. phantom 3D, merged 281 — very strong config. A 2D, averaged 432 1 strong config. A 2D, slice-merged 1.73 × 103 10 strong config. A 2.5D, direction-merged 9.85 × 103 10 strong config. A 2.5D, merged 3.94 × 104 100 strong config. B 2D, averaged 107 1 strong config. B 2D, slice-merged 427 1 strong config. B 2.5D, direction-merged 2.4 × 103 10 strong config. B 2.5D, merged 9.6 × 103 20 strong config. C 3D, averaged 3.18 × 103 10 strong config. C 3D, merged 4.13 × 104 100 strong config. D 3D, averaged 3.29 × 103 10 strong config. D 3D, merged 4.28 × 104 200 strong config. E 3D, averaged 4 × 103 10 moderate config. E 3D, merged 5.19 × 104 200 strong

Table 3.114 — Reference values for the grey level non-uniformity feature. CHAPTER 3. IMAGE FEATURES 90

3.7.10 Normalised grey level non-uniformity OVBL

This is a normalised version of the grey level non-uniformity feature. It is deﬁned as:

Ng 1 X F = r2 rlm.glnu.norm N 2 i. s i=1

data aggr. method value tol. consensus

dig. phantom 2D, averaged 0.46 — very strong dig. phantom 2D, slice-merged 0.456 — strong dig. phantom 2.5D, direction-merged 0.413 — strong dig. phantom 2.5D, merged 0.412 — strong dig. phantom 3D, averaged 0.43 — very strong dig. phantom 3D, merged 0.43 — very strong config. A 2D, averaged 0.128 0.003 strong config. A 2D, slice-merged 0.128 0.003 strong config. A 2.5D, direction-merged 0.126 0.003 strong config. A 2.5D, merged 0.126 0.003 strong config. B 2D, averaged 0.145 0.001 strong config. B 2D, slice-merged 0.145 0.001 strong config. B 2.5D, direction-merged 0.137 0.001 strong config. B 2.5D, merged 0.137 0.001 strong config. C 3D, averaged 0.102 0.003 strong config. C 3D, merged 0.102 0.003 very strong config. D 3D, averaged 0.133 0.002 strong config. D 3D, merged 0.134 0.002 strong config. E 3D, averaged 0.135 0.003 strong config. E 3D, merged 0.135 0.003 strong

Table 3.115 — Reference values for the normalised grey level non-uniformity feature.

3.7.11 Run length non-uniformity W92Y

This features assesses the distribution of runs over the run lengths 30. The feature value is low when runs are equally distributed along run lengths. It is deﬁned as:

N 1 Xr F = r2 rlm.rlnu N .j s j=1

data aggr. method value tol. consensus

dig. phantom 2D, averaged 6.12 — very strong dig. phantom 2D, slice-merged 21.6 — strong dig. phantom 2.5D, direction-merged 22.3 — strong dig. phantom 2.5D, merged 83.2 — strong CHAPTER 3. IMAGE FEATURES 91

dig. phantom 3D, averaged 26.9 — very strong dig. phantom 3D, merged 328 — very strong config. A 2D, averaged 1.65 × 103 10 strong config. A 2D, slice-merged 6.6 × 103 30 strong config. A 2.5D, direction-merged 4.27 × 104 200 strong config. A 2.5D, merged 1.71 × 105 1 × 103 strong config. B 2D, averaged 365 3 strong config. B 2D, slice-merged 1.46 × 103 10 strong config. B 2.5D, direction-merged 9.38 × 103 70 strong config. B 2.5D, merged 3.75 × 104 300 strong config. C 3D, averaged 1.8 × 104 500 strong config. C 3D, merged 2.34 × 105 6 × 103 strong config. D 3D, averaged 1.24 × 104 200 strong config. D 3D, merged 1.6 × 105 3 × 103 strong config. E 3D, averaged 1.66 × 104 300 strong config. E 3D, merged 2.15 × 105 4 × 103 strong

Table 3.116 — Reference values for the run length non-uniformity feature.

3.7.12 Normalised run length non-uniformity IC23

This is normalised version of the run length non-uniformity feature. It is deﬁned as:

N 1 Xr F = r2 rlm.rlnu.norm N 2 .j s j=1

data aggr. method value tol. consensus

dig. phantom 2D, averaged 0.492 — very strong dig. phantom 2D, slice-merged 0.441 — strong dig. phantom 2.5D, direction-merged 0.461 — strong dig. phantom 2.5D, merged 0.445 — strong dig. phantom 3D, averaged 0.513 — very strong dig. phantom 3D, merged 0.501 — very strong config. A 2D, averaged 0.579 0.003 strong config. A 2D, slice-merged 0.579 0.003 strong config. A 2.5D, direction-merged 0.548 0.003 strong config. A 2.5D, merged 0.548 0.003 strong config. B 2D, averaged 0.578 0.001 strong config. B 2D, slice-merged 0.578 0.001 strong config. B 2.5D, direction-merged 0.533 0.001 strong config. B 2.5D, merged 0.534 0.001 strong config. C 3D, averaged 0.574 0.004 strong config. C 3D, merged 0.575 0.004 strong config. D 3D, averaged 0.5 0.001 strong config. D 3D, merged 0.501 0.001 strong CHAPTER 3. IMAGE FEATURES 92

conﬁg. E 3D, averaged 0.559 0.001 moderate conﬁg. E 3D, merged 0.56 0.001 strong

Table 3.117 — Reference values for the normalised run length non-uniformity feature.

3.7.13 Run percentage 9ZK5

This feature measures the fraction of the number of realised runs and the maximum number of potential runs 30. Strongly linear or highly uniform ROI volumes produce a low run percentage. It is deﬁned as:

Ns Frlm.r.perc = Nv

As noted before, when this feature is calculated using a merged GLRLM, Nv should be the sum of the number of voxels of the underlying matrices to allow proper normalisation.

data aggr. method value tol. consensus

dig. phantom 2D, averaged 0.627 — very strong dig. phantom 2D, slice-merged 0.627 — strong dig. phantom 2.5D, direction-merged 0.632 — strong dig. phantom 2.5D, merged 0.632 — strong dig. phantom 3D, averaged 0.68 — very strong dig. phantom 3D, merged 0.68 — very strong config. A 2D, averaged 0.704 0.003 strong config. A 2D, slice-merged 0.704 0.003 strong config. A 2.5D, direction-merged 0.68 0.003 strong config. A 2.5D, merged 0.68 0.003 strong config. B 2D, averaged 0.681 0.002 strong config. B 2D, slice-merged 0.681 0.002 strong config. B 2.5D, direction-merged 0.642 0.002 strong config. B 2.5D, merged 0.642 0.002 strong config. C 3D, averaged 0.679 0.003 strong config. C 3D, merged 0.679 0.003 strong config. D 3D, averaged 0.554 0.005 strong config. D 3D, merged 0.554 0.005 strong config. E 3D, averaged 0.664 0.003 moderate config. E 3D, merged 0.664 0.003 strong

Table 3.118 — Reference values for the run percentage feature.

3.7.14 Grey level variance 8CE5

This feature estimates the variance in runs over the grey levels. Let pij = rij/Ns be the joint probability estimate for ﬁnding discretised grey level i with run length j. Grey level variance is CHAPTER 3. IMAGE FEATURES 93 then deﬁned as:

Ng Nr X X 2 Frlm.gl.var = (i − µ) pij i=1 j=1

PNg PNr Here, µ = i=1 j=1 i pij.

data aggr. method value tol. consensus

dig. phantom 2D, averaged 3.35 — very strong dig. phantom 2D, slice-merged 3.37 — strong dig. phantom 2.5D, direction-merged 3.58 — strong dig. phantom 2.5D, merged 3.59 — strong dig. phantom 3D, averaged 3.46 — very strong dig. phantom 3D, merged 3.48 — very strong config. A 2D, averaged 33.7 0.6 strong config. A 2D, slice-merged 33.7 0.6 strong config. A 2.5D, direction-merged 29.1 0.6 strong config. A 2.5D, merged 29.1 0.6 strong config. B 2D, averaged 28.3 0.3 strong config. B 2D, slice-merged 28.3 0.3 strong config. B 2.5D, direction-merged 25.7 0.2 strong config. B 2.5D, merged 25.7 0.2 strong config. C 3D, averaged 101 3 strong config. C 3D, merged 101 3 very strong config. D 3D, averaged 31.5 0.4 strong config. D 3D, merged 31.4 0.4 strong config. E 3D, averaged 39.8 0.9 moderate config. E 3D, merged 39.7 0.9 strong

Table 3.119 — Reference values for the grey level variance feature.

3.7.15 Run length variance SXLW

This feature estimates the variance in runs over the run lengths. As before let pij = rij/Ns. The feature is deﬁned as:

Ng Nr X X 2 Frlm.rl.var = (j − µ) pij i=1 j=1

PNg PNr Mean run length is deﬁned as µ = i=1 j=1 j pij.

data aggr. method value tol. consensus

dig. phantom 2D, averaged 0.761 — very strong dig. phantom 2D, slice-merged 0.778 — strong dig. phantom 2.5D, direction-merged 0.758 — strong dig. phantom 2.5D, merged 0.767 — strong CHAPTER 3. IMAGE FEATURES 94

dig. phantom 3D, averaged 0.574 — very strong dig. phantom 3D, merged 0.598 — very strong config. A 2D, averaged 0.828 0.008 strong config. A 2D, slice-merged 0.826 0.008 strong config. A 2.5D, direction-merged 0.916 0.011 strong config. A 2.5D, merged 0.914 0.011 strong config. B 2D, averaged 1.22 0.03 strong config. B 2D, slice-merged 1.21 0.03 strong config. B 2.5D, direction-merged 1.39 0.03 strong config. B 2.5D, merged 1.39 0.03 strong config. C 3D, averaged 1.12 0.02 strong config. C 3D, merged 1.11 0.02 strong config. D 3D, averaged 3.35 0.14 strong config. D 3D, merged 3.29 0.13 strong config. E 3D, averaged 1.26 0.05 strong config. E 3D, merged 1.25 0.05 strong

Table 3.120 — Reference values for the run length variance feature.

3.7.16 Run entropy HJ9O

3 Run entropy was investigated by Albregtsen et al. . Again, let pij = rij/Ns. The entropy is then deﬁned as:

Ng N X Xr Frlm.rl.entr = − pij log2 pij i=1 j=1

data aggr. method value tol. consensus

dig. phantom 2D, averaged 2.17 — very strong dig. phantom 2D, slice-merged 2.57 — strong dig. phantom 2.5D, direction-merged 2.52 — strong dig. phantom 2.5D, merged 2.76 — strong dig. phantom 3D, averaged 2.43 — very strong dig. phantom 3D, merged 2.62 — very strong config. A 2D, averaged 4.73 0.02 strong config. A 2D, slice-merged 4.76 0.02 strong config. A 2.5D, direction-merged 4.87 0.01 strong config. A 2.5D, merged 4.87 0.01 strong config. B 2D, averaged 4.53 0.02 strong config. B 2D, slice-merged 4.58 0.01 strong config. B 2.5D, direction-merged 4.84 0.01 strong config. B 2.5D, merged 4.84 0.01 strong config. C 3D, averaged 5.35 0.03 strong config. C 3D, merged 5.35 0.03 very strong config. D 3D, averaged 5.08 0.02 strong config. D 3D, merged 5.08 0.02 strong CHAPTER 3. IMAGE FEATURES 95

conﬁg. E 3D, averaged 4.87 0.03 strong conﬁg. E 3D, merged 4.87 0.03 strong

Table 3.121 — Reference values for the run entropy feature. CHAPTER 3. IMAGE FEATURES 96

3.8 Grey level size zone based features 9SAK

The grey level size zone matrix (GLSZM) counts the number of groups (or zones) of linked voxels 71. Voxels are linked if the neighbouring voxel has an identical discretised grey level. Whether a voxel classiﬁes as a neighbour depends on its connectedness. In a 3D approach to texture analysis we consider 26-connectedness, which indicates that a center voxel is linked to all of the 26 neighbouring voxels with the same grey level. In the 2 dimensional approach, 8-connectedness is used. A potential issue for the 2D approach is that voxels which may otherwise be considered to belong to the same zone by linking across slices, are now two or more separate zones within the slice plane. Whether this issue negatively aﬀects predictive performance of GLSZM-based features or their reproducibility has not been determined.

Let M be the Ng × Nz grey level size zone matrix, where Ng is the number of discretised grey levels present in the ROI intensity mask and Nz the maximum zone size of any group of linked voxels. Element sij of M is then the number of zones with discretised grey level i and size j. PNg PNz Furthermore, let Nv be the number of voxels in the intensity mask and Ns = i=1 j=1 sij be the PNz total number of zones. Marginal sums can likewise be deﬁned. Let si. = j=1 sij be the number PNg of zones with discretised grey level i, regardless of size. Likewise, let s.j = i=1 sij be the number of zones with size j, regardless of grey level. A two dimensional example is shown in Table 3.122.

Aggregating features Three methods can be used to aggregate GLSZMs and arrive at a single feature value. A schematic example is shown in Figure 3.4. A feature may be aggregated as follows:

1. Features are computed from 2D matrices and averaged over slices (8QNN).

2. The feature is computed from a single matrix after merging all 2D matrices (62GR).

3. The feature is computed from a 3D matrix (KOBO).

Method 2 involves merging GLSZMs by summing the number of zones sij over the GLSZM for the diﬀerent slices. Note that when matrices are merged, Nv should likewise be summed to retain consistency. Feature values may dependent strongly on the aggregation method.

Distances The default neighbourhood for GLSZM is constructed using Chebyshev distance δ = 1. Man- hattan or Euclidean norms may also be used to construct a neighbourhood, and both lead to a 6-connected (3D) and 4-connected (2D) neighbourhoods. Larger distances are also technically possible, but will occasionally cause separate zones with the same intensity to be considered as belonging to the same zone. Using diﬀerent neighbourhoods for determining voxel linkage is non- standard use, and we caution against it due to potential reproducibility issues.

Note on feature references GLSZM feature deﬁnitions are based on the deﬁnitions of GLRLM features 71. Hence, references may be found in the section on GLRLM (3.7). CHAPTER 3. IMAGE FEATURES 97

Zone size j 1 2 3 4 5 1 2 2 3 1 2 1 0 0 0 1 2 3 3 2 0 0 0 0 1 i 4 2 4 1 3 1 0 1 0 0 4 1 2 3 4 1 1 0 0 0

(a) Grey levels (b) Grey level size zone matrix Table 3.122 — Original image with grey levels (a); and corresponding grey level size zone matrix (GLSZM) under 8-connectedness (b). Element s(i, j) of the GLSZM indicates the number of times a zone of j linked pixels and grey level i occurs within the image.

(a) 2D: by slice, without merging (b) 2.5D: by slice, with merging

Figure 3.4 — Approaches to calculating grey level size zone matrix-based features. Mk are texture matrices calculated for slice k (if applicable), and fk is the corresponding feature value. In (b) the matrices from the diﬀerent slices are merged prior to feature calculation.

3.8.1 Small zone emphasis 5QRC

This feature emphasises small zones. It is deﬁned as:

Nz 1 X s.j F = szm.sze N j2 s j=1

data aggr. method value tol. consensus

dig. phantom 2D 0.363 — strong dig. phantom 2.5D 0.368 — strong dig. phantom 3D 0.255 — very strong config. A 2D 0.688 0.003 strong config. A 2.5D 0.68 0.003 strong config. B 2D 0.745 0.003 strong config. B 2.5D 0.741 0.003 strong config. C 3D 0.695 0.001 strong config. D 3D 0.637 0.005 strong CHAPTER 3. IMAGE FEATURES 98

conﬁg. E 3D 0.676 0.003 strong

Table 3.123 — Reference values for the small zone emphasis feature.

3.8.2 Large zone emphasis 48P8

This feature emphasises large zones. It is deﬁned as:

N 1 Xz F = j2s szm.lze N .j s j=1

data aggr. method value tol. consensus

dig. phantom 2D 43.9 — strong dig. phantom 2.5D 34.2 — strong dig. phantom 3D 550 — very strong config. A 2D 625 9 strong config. A 2.5D 675 8 strong config. B 2D 439 8 strong config. B 2.5D 444 8 strong config. C 3D 3.89 × 104 900 strong config. D 3D 9.91 × 104 2.8 × 103 strong config. E 3D 5.86 × 104 800 strong

Table 3.124 — Reference values for the large zone emphasis feature.

3.8.3 Low grey level zone emphasis XMSY

This feature is a grey level analogue to small zone emphasis. Instead of small zone sizes, low grey levels are emphasised. The feature is deﬁned as:

Ng 1 X si. F = szm.lgze N i2 s i=1

data aggr. method value tol. consensus

dig. phantom 2D 0.371 — strong dig. phantom 2.5D 0.368 — strong dig. phantom 3D 0.253 — very strong config. A 2D 0.0368 0.0005 strong config. A 2.5D 0.0291 0.0005 strong config. B 2D 0.0475 0.001 strong config. B 2.5D 0.0387 0.001 strong config. C 3D 0.00235 6 × 10−5 strong config. D 3D 0.0409 0.0005 strong config. E 3D 0.034 0.0004 strong CHAPTER 3. IMAGE FEATURES 99

Table 3.125 — Reference values for the low grey level emphasis feature.

3.8.4 High grey level zone emphasis 5GN9

The high grey level zone emphasis feature is a grey level analogue to large zone emphasis. The feature emphasises high grey levels, and is deﬁned as:

Ng 1 X F = i2s szm.hgze N i. s i=1

data aggr. method value tol. consensus

dig. phantom 2D 16.4 — strong dig. phantom 2.5D 16.2 — strong dig. phantom 3D 15.6 — very strong config. A 2D 363 3 strong config. A 2.5D 370 3 strong config. B 2D 284 11 strong config. B 2.5D 284 11 strong config. C 3D 971 7 strong config. D 3D 188 10 strong config. E 3D 286 6 strong

Table 3.126 — Reference values for the high grey level emphasis feature.

3.8.5 Small zone low grey level emphasis 5RAI

This feature emphasises zone counts within the upper left quadrant of the GLSZM, where small zone sizes and low grey levels are located. It is deﬁned as:

Ng Nz 1 X X sij F = szm.szlge N i2j2 s i=1 j=1

data aggr. method value tol. consensus

dig. phantom 2D 0.0259 — strong dig. phantom 2.5D 0.0295 — strong dig. phantom 3D 0.0256 — very strong config. A 2D 0.0298 0.0005 strong config. A 2.5D 0.0237 0.0005 strong config. B 2D 0.0415 0.0008 strong config. B 2.5D 0.0335 0.0009 strong config. C 3D 0.0016 4 × 10−5 strong config. D 3D 0.0248 0.0004 strong config. E 3D 0.0224 0.0004 strong CHAPTER 3. IMAGE FEATURES 100

Table 3.127 — Reference values for the small zone low grey level emphasis feature.

3.8.6 Small zone high grey level emphasis HW1V

This feature emphasises zone counts in the lower left quadrant of the GLSZM, where small zone sizes and high grey levels are located. The feature is deﬁned as:

Ng Nz 2 1 X X i sij F = szm.szhge N j2 s i=1 j=1

data aggr. method value tol. consensus

dig. phantom 2D 10.3 — strong dig. phantom 2.5D 9.87 — strong dig. phantom 3D 2.76 — very strong config. A 2D 226 1 strong config. A 2.5D 229 1 strong config. B 2D 190 7 strong config. B 2.5D 190 7 strong config. C 3D 657 4 strong config. D 3D 117 7 strong config. E 3D 186 4 strong

Table 3.128 — Reference values for the small zone high grey level emphasis feature.

3.8.7 Large zone low grey level emphasis YH51

This feature emphasises zone counts in the upper right quadrant of the GLSZM, where large zone sizes and low grey levels are located. The feature is deﬁned as:

Ng Nz 2 1 X X j sij F = szm.lzlge N i2 s i=1 j=1

data aggr. method value tol. consensus

dig. phantom 2D 40.4 — strong dig. phantom 2.5D 30.6 — strong dig. phantom 3D 503 — very strong config. A 2D 1.35 0.03 strong config. A 2.5D 1.44 0.02 strong config. B 2D 1.15 0.04 strong config. B 2.5D 1.16 0.04 strong config. C 3D 21.6 0.5 strong config. D 3D 241 14 strong config. E 3D 105 4 strong CHAPTER 3. IMAGE FEATURES 101

Table 3.129 — Reference values for the large zone low grey level emphasis feature.

3.8.8 Large zone high grey level emphasis J17V

This feature emphasises zone counts in the lower right quadrant of the GLSZM, where large zone sizes and high grey levels are located. The feature is deﬁned as:

Ng N 1 X Xz F = i2j2s szm.lzhge N ij s i=1 j=1

data aggr. method value tol. consensus

dig. phantom 2D 113 — strong dig. phantom 2.5D 107 — strong dig. phantom 3D 1.49 × 103 — very strong config. A 2D 3.16 × 105 5 × 103 strong config. A 2.5D 3.38 × 105 5 × 103 strong config. B 2D 1.81 × 105 3 × 103 strong config. B 2.5D 1.81 × 105 3 × 103 strong config. C 3D 7.07 × 107 1.5 × 106 strong config. D 3D 4.14 × 107 3 × 105 strong config. E 3D 3.36 × 107 3 × 105 strong

Table 3.130 — Reference values for the large zone high grey level emphasis feature.

3.8.9 Grey level non-uniformity JNSA

This feature assesses the distribution of zone counts over the grey values. The feature value is low when zone counts are equally distributed along grey levels. The feature is deﬁned as:

Ng 1 X F = s2 szm.glnu N i. s i=1

data aggr. method value tol. consensus

dig. phantom 2D 1.41 — strong dig. phantom 2.5D 5.44 — strong dig. phantom 3D 1.4 — very strong config. A 2D 82.2 0.1 strong config. A 2.5D 1.8 × 103 10 strong config. B 2D 20.5 0.1 strong config. B 2.5D 437 3 strong config. C 3D 195 6 strong config. D 3D 212 6 very strong config. E 3D 231 6 strong CHAPTER 3. IMAGE FEATURES 102

Table 3.131 — Reference values for the grey level non-uniformity feature.

3.8.10 Normalised grey level non-uniformity Y1RO

This is a normalised version of the grey level non-uniformity feature. It is deﬁned as:

Ng 1 X F = s2 szm.glnu.norm N 2 i. s i=1

data aggr. method value tol. consensus

dig. phantom 2D 0.323 — strong dig. phantom 2.5D 0.302 — strong dig. phantom 3D 0.28 — very strong config. A 2D 0.0728 0.0014 strong config. A 2.5D 0.0622 0.0007 strong config. B 2D 0.0789 0.001 strong config. B 2.5D 0.0613 0.0005 strong config. C 3D 0.0286 0.0003 strong config. D 3D 0.0491 0.0008 strong config. E 3D 0.0414 0.0003 strong

Table 3.132 — Reference values for the normalised grey level non-uniformity feature.

3.8.11 Zone size non-uniformity 4JP3

This features assesses the distribution of zone counts over the diﬀerent zone sizes. Zone size non-uniformity is low when zone counts are equally distributed along zone sizes. It is deﬁned as:

N 1 Xz F = s2 szm.zsnu N .j s j=1

data aggr. method value tol. consensus

dig. phantom 2D 1.49 — strong dig. phantom 2.5D 3.44 — strong dig. phantom 3D 1 — very strong config. A 2D 479 4 strong config. A 2.5D 1.24 × 104 100 strong config. B 2D 140 3 strong config. B 2.5D 3.63 × 103 70 strong config. C 3D 3.04 × 103 100 strong config. D 3D 1.63 × 103 10 strong config. E 3D 2.37 × 103 40 strong

Table 3.133 — Reference values for the zone size non-uniformity feature. CHAPTER 3. IMAGE FEATURES 103

3.8.12 Normalised zone size non-uniformity VB3A

This is a normalised version of zone size non-uniformity. It is deﬁned as:

N 1 Xz F = s2 szm.zsnu.norm N 2 .j s i=1

data aggr. method value tol. consensus

dig. phantom 2D 0.333 — strong dig. phantom 2.5D 0.191 — strong dig. phantom 3D 0.2 — very strong config. A 2D 0.44 0.004 strong config. A 2.5D 0.427 0.004 strong config. B 2D 0.521 0.004 strong config. B 2.5D 0.509 0.004 strong config. C 3D 0.447 0.001 strong config. D 3D 0.377 0.006 strong config. E 3D 0.424 0.004 strong

Table 3.134 — Reference values for the normalised zone size non-uniformity feature.

3.8.13 Zone percentage P30P

This feature measures the fraction of the number of realised zones and the maximum number of potential zones. Highly uniform ROIs produce a low zone percentage. It is deﬁned as:

Ns Fszm.z.perc = Nv

data aggr. method value tol. consensus

dig. phantom 2D 0.24 — strong dig. phantom 2.5D 0.243 — strong dig. phantom 3D 0.0676 — very strong config. A 2D 0.3 0.003 strong config. A 2.5D 0.253 0.004 strong config. B 2D 0.324 0.001 strong config. B 2.5D 0.26 0.002 strong config. C 3D 0.148 0.003 very strong config. D 3D 0.0972 0.0007 strong config. E 3D 0.126 0.001 strong

Table 3.135 — Reference values for the zone percentage feature. CHAPTER 3. IMAGE FEATURES 104

3.8.14 Grey level variance BYLV

This feature estimates the variance in zone counts over the grey levels. Let pij = sij/Ns be the joint probability estimate for ﬁnding zones with discretised grey level i and size j. The feature is then deﬁned as:

Ng Nz X X 2 Fszm.gl.var = (i − µ) pij i=1 j=1

PNg PNz Here, µ = i=1 j=1 i pij.

data aggr. method value tol. consensus

dig. phantom 2D 3.97 — strong dig. phantom 2.5D 3.92 — strong dig. phantom 3D 2.64 — very strong config. A 2D 42.7 0.7 strong config. A 2.5D 47.9 0.4 strong config. B 2D 36.1 0.3 strong config. B 2.5D 41 0.7 strong config. C 3D 106 1 strong config. D 3D 32.7 1.6 strong config. E 3D 50.8 0.9 strong

Table 3.136 — Reference values for the grey level variance feature.

3.8.15 Zone size variance 3NSA

This feature estimates the variance in zone counts over the diﬀerent zone sizes. As before let pij = sij/Ns. The feature is deﬁned as:

Ng Nz X X 2 Fszm.zs.var = (j − µ) pij i=1 j=1

PNg PNz Mean zone size is deﬁned as µ = i=1 j=1 j pij.

data aggr. method value tol. consensus

dig. phantom 2D 21 — strong dig. phantom 2.5D 17.3 — strong dig. phantom 3D 331 — very strong config. A 2D 609 9 strong config. A 2.5D 660 8 strong config. B 2D 423 8 strong config. B 2.5D 429 8 strong config. C 3D 3.89 × 104 900 strong config. D 3D 9.9 × 104 2.8 × 103 strong config. E 3D 5.85 × 104 800 strong CHAPTER 3. IMAGE FEATURES 105

Table 3.137 — Reference values for the zone size variance feature.

3.8.16 Zone size entropy GU8N

Let pij = sij/Ns. Zone size entropy is then deﬁned as:

Ng N X Xz Fszm.zs.entr = − pij log2 pij i=1 j=1

data aggr. method value tol. consensus

dig. phantom 2D 1.93 — strong dig. phantom 2.5D 3.08 — strong dig. phantom 3D 2.32 — very strong config. A 2D 5.92 0.02 strong config. A 2.5D 6.39 0.01 strong config. B 2D 5.29 0.01 strong config. B 2.5D 5.98 0.02 strong config. C 3D 7 0.01 strong config. D 3D 6.52 0.01 strong config. E 3D 6.57 0.01 strong

Table 3.138 — Reference values for the zone size entropy feature. CHAPTER 3. IMAGE FEATURES 106

3.9 Grey level distance zone based features VMDZ

The grey level distance zone matrix (GLDZM) counts the number of groups (or zones) of linked voxels which share a specific discretised grey level value and possess the same distance to ROI edge 71. The GLDZM thus captures the relation between location and grey level. Two maps are required to calculate the GLDZM. The first is a grey level zone map, which is identical to the one created for the grey level size zone matrix (GLSZM), see Section 3.8. The second is a distance map, which will be described in detail later. As with GSLZM, neighbouring voxels are linked if they share the same grey level value. Whether a voxel classifies as a neighbour depends on its connectedness. We consider 26-connectedness for a 3D approach and 8-connectedness in the 2D approach. The distance to the ROI edge is defined according to 6 and 4-connectedness for 3D and 2D, respectively. Because of the connectedness definition used, the distance of a voxel to the outer border is equal to the minimum number edges of neighbouring voxels that need to be crossed to reach the ROI edge. The distance for a linked group of voxels with the same grey value is equal to the minimum distance for the respective voxels in the distance map. Our definition deviates from the original by Thibault et al. 71 . The original was defined in a rectangular 2D image, whereas ROIs are rarely rectangular cuboids. Approximating distance using Chamfer maps is then no longer a fast and easy solution. Determining distance iteratively in 6 or 4-connectedness is a relatively efficient solution, implemented as follows:

1. The ROI mask is morphologically eroded using the appropriate (6 or 4-connected) structure element.

2. All eroded ROI voxels are updated in the distance map by adding 1.

3. The above steps are performed iteratively until the ROI mask is empty.

A second diﬀerence with the original deﬁnition is that the lowest possible distance is 1 instead of 0 for voxels directly on the ROI edge. This prevents division by 0 for some features.

Let M be the Ng × Nd grey level size zone matrix, where Ng is the number of discretised grey levels present in the ROI intensity mask and Nd the largest distance of any zone. Element dij = d(i, j) of M is then number of zones with discretised grey level i and distance j. Furthermore, PNg PNd let Nv be the number of voxels and Ns = i=1 j=1 dij be the total zone count. Marginal sums PNd can likewise be deﬁned. Let di. = j=1 dij be the number of zones with discretised grey level PNg i, regardless of distance. Likewise, let d.j = i=1 dij be the number of zones with distance j, regardless of grey level. A two dimensional example is shown in Table 3.139.

Morphological and intensity masks. The GLDZM is special in that it uses both ROI masks. The distance map is determined using the morphological ROI mask, whereas the intensity mask is used for determining the zones, as with the GLSZM.

Aggregating features Three methods can be used to aggregate GLDZMs and arrive at a single feature value. A schematic example was previously shown in Figure 3.4. A feature may be aggregated as follows:

1. Features are computed from 2D matrices and averaged over slices (8QNN).

2. The feature is computed from a single matrix after merging all 2D matrices (62GR). CHAPTER 3. IMAGE FEATURES 107

j 1 2 1 2 2 3 1 1 1 1 1 3 0 1 2 3 3 1 2 2 1 2 2 0 i 4 2 4 1 1 2 2 1 3 2 0 4 1 2 3 1 1 1 1 4 1 1

(a) Grey levels (b) Distance (c) Grey level map distance zone matrix Table 3.139 — Original image with grey levels (a); corresponding distance map for distance to border (b); and corresponding grey level distance zone matrix (GLDZM) under 4-connectedness (c). Element d(i, j) of the GLDZM indicates the number of times a zone with grey level i and a minimum distance to border j occurs within the image.

3. The feature is computed from a 3D matrix (KOBO).

Method 2 involves merging GLDZMs by summing the number of zones dij over the GLDZM for the diﬀerent slices. Note that when matrices are merged, Nv should likewise be summed to retain consistency. Feature values may dependent strongly on the aggregation method.

Distances In addition to the use of different distance norms to determine voxel linkage, as described in section 3.8, different distance norms may be used to determine distance of zones to the boundary. The default is to use the Manhattan norm which allows for a computationally efficient implementation, as described above. A similar implementation is possible using the Chebyshev norm, as it merely changes connectedness of the structure element. Implementations using an Euclidean distance norm are less efficient as this demands searching for the nearest non-ROI voxel for each of the Nv voxels in the ROI. An added issue is that Euclidean norms may lead to a wide range of different distances j that require rounding before constructing the grey level distance zone matrix M. Using different distance norms is non-standard use, and we caution against it due to potential reproducibility issues.

Note on feature references GLDZM feature deﬁnitions are based on the deﬁnitions of GLRLM features 71. Hence, references may be found in the section on GLRLM (3.7).

3.9.1 Small distance emphasis 0GBI

This feature emphasises small distances. It is deﬁned as:

Nd 1 X d.j F = dzm.sde N j2 s j=1

data aggr. method value tol. consensus

dig. phantom 2D 0.946 — strong dig. phantom 2.5D 0.917 — moderate CHAPTER 3. IMAGE FEATURES 108

dig. phantom 3D 1 — very strong config. A 2D 0.192 0.006 strong config. A 2.5D 0.168 0.005 strong config. B 2D 0.36 0.005 strong config. B 2.5D 0.329 0.004 strong config. C 3D 0.531 0.006 strong config. D 3D 0.579 0.004 strong config. E 3D 0.527 0.004 moderate

Table 3.140 — Reference values for the small distance emphasis feature.

3.9.2 Large distance emphasis MB4I

This feature emphasises large distances. It is deﬁned as:

N 1 Xd F = j2d dzm.lde N .j s j=1

data aggr. method value tol. consensus

dig. phantom 2D 1.21 — strong dig. phantom 2.5D 1.33 — moderate dig. phantom 3D 1 — very strong config. A 2D 161 1 moderate config. A 2.5D 178 1 moderate config. B 2D 31.6 0.2 moderate config. B 2.5D 34.3 0.2 moderate config. C 3D 11 0.3 strong config. D 3D 10.3 0.1 strong config. E 3D 12.6 0.1 moderate

Table 3.141 — Reference values for the large distance emphasis feature.

3.9.3 Low grey level zone emphasis S1RA

This feature is a grey level analogue to small distance emphasis. Instead of small zone distances, low grey levels are emphasised. The feature is deﬁned as:

Ng 1 X di. F = dzm.lgze N i2 s i=1

data aggr. method value tol. consensus

dig. phantom 2D 0.371 — strong dig. phantom 2.5D 0.368 — moderate dig. phantom 3D 0.253 — very strong CHAPTER 3. IMAGE FEATURES 109

config. A 2D 0.0368 0.0005 strong config. A 2.5D 0.0291 0.0005 strong config. B 2D 0.0475 0.001 strong config. B 2.5D 0.0387 0.001 strong config. C 3D 0.00235 6 × 10−5 strong config. D 3D 0.0409 0.0005 strong config. E 3D 0.034 0.0004 moderate

Table 3.142 — Reference values for the low grey level emphasis feature.

3.9.4 High grey level zone emphasis K26C

The high grey level zone emphasis feature is a grey level analogue to large distance emphasis. The feature emphasises high grey levels, and is deﬁned as:

Ng 1 X F = i2d dzm.hgze N i. s i=1

data aggr. method value tol. consensus

dig. phantom 2D 16.4 — strong dig. phantom 2.5D 16.2 — moderate dig. phantom 3D 15.6 — very strong config. A 2D 363 3 strong config. A 2.5D 370 3 strong config. B 2D 284 11 strong config. B 2.5D 284 11 strong config. C 3D 971 7 strong config. D 3D 188 10 strong config. E 3D 286 6 strong

Table 3.143 — Reference values for the high grey level emphasis feature.

3.9.5 Small distance low grey level emphasis RUVG

This feature emphasises runs in the upper left quadrant of the GLDZM, where small zone distances and low grey levels are located. It is deﬁned as:

Ng Nd 1 X X dij F = dzm.sdlge N i2j2 s i=1 j=1

data aggr. method value tol. consensus

dig. phantom 2D 0.367 — strong dig. phantom 2.5D 0.362 — moderate dig. phantom 3D 0.253 — very strong CHAPTER 3. IMAGE FEATURES 110

config. A 2D 0.00913 0.00023 strong config. A 2.5D 0.00788 0.00022 strong config. B 2D 0.0192 0.0005 strong config. B 2.5D 0.0168 0.0005 strong config. C 3D 0.00149 4 × 10−5 strong config. D 3D 0.0302 0.0006 strong config. E 3D 0.0228 0.0003 moderate

Table 3.144 — Reference values for the small distance low grey level emphasis feature.

3.9.6 Small distance high grey level emphasis DKNJ

This feature emphasises runs in the lower left quadrant of the GLDZM, where small zone distances and high grey levels are located. Small distance high grey level emphasis is deﬁned as:

Ng Nd 2 1 X X i dij F = dzm.sdhge N j2 s i=1 j=1

data aggr. method value tol. consensus

dig. phantom 2D 15.2 — strong dig. phantom 2.5D 14.3 — moderate dig. phantom 3D 15.6 — very strong config. A 2D 60.1 3.3 strong config. A 2.5D 49.5 2.8 strong config. B 2D 95.7 5.5 strong config. B 2.5D 81.4 4.6 strong config. C 3D 476 11 strong config. D 3D 99.3 5.1 strong config. E 3D 136 4 moderate

Table 3.145 — Reference values for the small distance high grey level emphasis feature.

3.9.7 Large distance low grey level emphasis A7WM

This feature emphasises runs in the upper right quadrant of the GLDZM, where large zone distances and low grey levels are located. The feature is deﬁned as:

Ng Nd 2 1 X X j dij F = dzm.ldlge N i2 s i=1 j=1

data aggr. method value tol. consensus

dig. phantom 2D 0.386 — strong dig. phantom 2.5D 0.391 — moderate dig. phantom 3D 0.253 — very strong CHAPTER 3. IMAGE FEATURES 111

config. A 2D 2.96 0.02 moderate config. A 2.5D 2.31 0.01 moderate config. B 2D 0.934 0.018 moderate config. B 2.5D 0.748 0.017 moderate config. C 3D 0.0154 0.0005 strong config. D 3D 0.183 0.004 strong config. E 3D 0.179 0.004 moderate

Table 3.146 — Reference values for the large distance low grey level emphasis feature.

3.9.8 Large distance high grey level emphasis KLTH

This feature emphasises runs in the lower right quadrant of the GLDZM, where large zone distances and high grey levels are located. The large distance high grey level emphasis feature is deﬁned as:

Ng N 1 X Xd F = i2j2d dzm.ldhge N ij s i=1 j=1

data aggr. method value tol. consensus

dig. phantom 2D 21.3 — strong dig. phantom 2.5D 23.7 — moderate dig. phantom 3D 15.6 — very strong config. A 2D 7.01 × 104 100 moderate config. A 2.5D 7.95 × 104 100 moderate config. B 2D 1.06 × 104 300 strong config. B 2.5D 1.16 × 104 400 strong config. C 3D 1.34 × 104 200 strong config. D 3D 2.62 × 103 110 strong config. E 3D 4.85 × 103 60 moderate

Table 3.147 — Reference values for the large distance high grey level emphasis feature.

3.9.9 Grey level non-uniformity VFT7

This feature measures the distribution of zone counts over the grey values. Grey level non- uniformity is low when zone counts are equally distributed along grey levels. The feature is deﬁned as:

Ng 1 X F = d2 dzm.glnu N i. s i=1

data aggr. method value tol. consensus

dig. phantom 2D 1.41 — strong dig. phantom 2.5D 5.44 — moderate dig. phantom 3D 1.4 — very strong CHAPTER 3. IMAGE FEATURES 112

config. A 2D 82.2 0.1 strong config. A 2.5D 1.8 × 103 10 strong config. B 2D 20.5 0.1 strong config. B 2.5D 437 3 strong config. C 3D 195 6 strong config. D 3D 212 6 strong config. E 3D 231 6 moderate

Table 3.148 — Reference values for the grey level non-uniformity feature.

3.9.10 Normalised grey level non-uniformity 7HP3

This is a normalised version of the grey level non-uniformity feature. It is deﬁned as:

Ng 1 X F = d2 dzm.glnu.norm N 2 i. s i=1

data aggr. method value tol. consensus

dig. phantom 2D 0.323 — strong dig. phantom 2.5D 0.302 — moderate dig. phantom 3D 0.28 — very strong config. A 2D 0.0728 0.0014 strong config. A 2.5D 0.0622 0.0007 strong config. B 2D 0.0789 0.001 strong config. B 2.5D 0.0613 0.0005 strong config. C 3D 0.0286 0.0003 strong config. D 3D 0.0491 0.0008 strong config. E 3D 0.0414 0.0003 moderate

Table 3.149 — Reference values for the normalised grey level non-uniformity feature.

3.9.11 Zone distance non-uniformity V294

Zone distance non-uniformity measures the distribution of zone counts over the diﬀerent zone distances. Zone distance non-uniformity is low when zone counts are equally distributed along zone distances. It is deﬁned as:

N 1 Xd F = d2 dzm.zdnu N .j s j=1

data aggr. method value tol. consensus

dig. phantom 2D 3.79 — strong dig. phantom 2.5D 14.4 — moderate dig. phantom 3D 5 — very strong conﬁg. A 2D 64 0.4 moderate CHAPTER 3. IMAGE FEATURES 113

config. A 2.5D 1.57 × 103 10 strong config. B 2D 39.8 0.3 moderate config. B 2.5D 963 6 moderate config. C 3D 1.87 × 103 40 strong config. D 3D 1.37 × 103 20 strong config. E 3D 1.5 × 103 30 moderate

Table 3.150 — Reference values for the zone distance non-uniformity feature.

3.9.12 Normalised zone distance non-uniformity IATH

This is a normalised version of the zone distance non-uniformity feature. It is deﬁned as:

N 1 Xd F = d2 dzm.zdnu.norm N 2 .j s i=1

data aggr. method value tol. consensus

dig. phantom 2D 0.898 — strong dig. phantom 2.5D 0.802 — moderate dig. phantom 3D 1 — very strong config. A 2D 0.0716 0.0022 strong config. A 2.5D 0.0543 0.0014 strong config. B 2D 0.174 0.003 strong config. B 2.5D 0.135 0.001 strong config. C 3D 0.274 0.005 strong config. D 3D 0.317 0.004 strong config. E 3D 0.269 0.003 moderate

Table 3.151 — Reference values for the normalised zone distance non-uniformity feature.

3.9.13 Zone percentage VIWW

This feature measures the fraction of the number of realised zones and the maximum number of potential zones. Highly uniform ROIs produce a low zone percentage. It is deﬁned as:

Ns Fdzm.z.perc = Nv

data aggr. method value tol. consensus

dig. phantom 2D 0.24 — strong dig. phantom 2.5D 0.243 — moderate dig. phantom 3D 0.0676 — very strong config. A 2D 0.3 0.003 strong config. A 2.5D 0.253 0.004 moderate config. B 2D 0.324 0.001 strong CHAPTER 3. IMAGE FEATURES 114

config. B 2.5D 0.26 0.002 moderate config. C 3D 0.148 0.003 strong config. D 3D 0.0972 0.0007 strong config. E 3D 0.126 0.001 moderate

Table 3.152 — Reference values for the zone percentage feature.

3.9.14 Grey level variance QK93

This feature estimates the variance in zone counts over the grey levels. Let pij = dij/Ns be the joint probability estimate for ﬁnding zones with discretised grey level i at distance j. The feature is then deﬁned as:

Ng Nd X X 2 Fdzm.gl.var = (i − µ) pij i=1 j=1

PNg PNd Here, µ = i=1 j=1 i pij.

data aggr. method value tol. consensus

dig. phantom 2D 3.97 — strong dig. phantom 2.5D 3.92 — moderate dig. phantom 3D 2.64 — very strong config. A 2D 42.7 0.7 moderate config. A 2.5D 47.9 0.4 strong config. B 2D 36.1 0.3 moderate config. B 2.5D 41 0.7 strong config. C 3D 106 1 strong config. D 3D 32.7 1.6 strong config. E 3D 50.8 0.9 strong

Table 3.153 — Reference values for the grey level variance feature.

3.9.15 Zone distance variance 7WT1

This feature estimates the variance in zone counts for the diﬀerent zone distances. As before let pij = dij/Ns. The feature is deﬁned as:

Ng Nd X X 2 Fdzm.zd.var = (j − µ) pij i=1 j=1

PNg PNd Mean zone size is deﬁned as µ = i=1 j=1 j pij.

data aggr. method value tol. consensus

dig. phantom 2D 0.051 — strong dig. phantom 2.5D 0.0988 — moderate dig. phantom 3D 0 — very strong conﬁg. A 2D 69.4 0.1 moderate CHAPTER 3. IMAGE FEATURES 115

config. A 2.5D 78.9 0.1 moderate config. B 2D 13.5 0.1 moderate config. B 2.5D 15 0.1 moderate config. C 3D 4.6 0.06 strong config. D 3D 4.61 0.04 strong config. E 3D 5.56 0.05 strong

Table 3.154 — Reference values for the zone distance variance feature.

3.9.16 Zone distance entropy GBDU

Again, let pij = dij/Ns. Zone distance entropy is then deﬁned as:

Ng N X Xd Fdzm.zd.entr = − pij log2 pij i=1 j=1

data aggr. method value tol. consensus

dig. phantom 2D 1.73 — strong dig. phantom 2.5D 2 — moderate dig. phantom 3D 1.92 — very strong config. A 2D 8 0.04 strong config. A 2.5D 8.87 0.03 strong config. B 2D 6.47 0.03 strong config. B 2.5D 7.58 0.01 moderate config. C 3D 7.56 0.03 strong config. D 3D 6.61 0.03 strong config. E 3D 7.06 0.01 moderate

Table 3.155 — Reference values for the zone distance entropy feature. CHAPTER 3. IMAGE FEATURES 116

3.10 Neighbourhood grey tone diﬀerence based featuresIPET

Amadasun and King 5 introduced an alternative to the grey level co-occurrence matrix. The neighbourhood grey tone difference matrix (NGTDM) contains the sum of grey level differences of pixels/voxels with discretised grey level i and the average discretised grey level of neighbouring pixels/voxels within a Chebyshev distance δ. For 3D volumes, we can extend the original definition by Amadasun and King. Let Xd,k be the discretised grey level of a voxel at position k = (kx, ky, kz). Then the average grey level within a neighbourhood centred at (kx, ky, kz), but excluding (kx, ky, kz) itself is:

δ δ δ 1 X X X X = X (k +m , k +m , k +m ) k W d x x y y z z mz =−δ my =−δ mx=−δ

(mx, my, mz) 6= (0, 0, 0)

W = (2δ + 1)3 − 1 is the size of the 3D neighbourhood. For 2D W = (2δ + 1)2 − 1, and averages are not calculated between diﬀerent slices. Neighbourhood grey tone diﬀerence si for discretised grey level i is then:

N Xv si = |i − Xk| [Xd(k) = i and k has a valid neighbourhood] k

Here, [...] is an Iverson bracket, which is 1 if the conditions that the grey level Xd,k of voxel k is equal to i and the voxel has a valid neighbourhood are both true; it is 0 otherwise. Nv is the number of voxels in the ROI intensity mask. A 2D example is shown in Table 3.156. A distance of δ = 1 is used in this example, leading to 8 neighbouring pixels. Entry s1 = 0 because there are no valid pixels with grey level 1. Two pixels have grey level 2. The average value of their neighbours are 19/8 and 21/8. Thus s2 = |2 − 19/8| + |2 − 21/8| = 1. Similarly s3 = |3 − 19/8| = 0.625 and s4 = |4 − 17/8| = 1.825. We deviate from the original deﬁnition by Amadasun and King 5 as we do not demand that valid neighbourhoods are completely inside the ROI. In an irregular ROI mask, valid neighbourhoods may simply not exist for a distance δ. Instead, we consider a valid neighbourhood to exist if there is at least one neighbouring voxel included in the ROI mask. The average grey level for voxel k within a valid neighbourhood is then:

δ δ δ 1 X X X Xk = Xd(k + m)[m 6= 0 and k + m in ROI] Wk mz =−δ my =−δ mx=−δ

The neighbourhood size Wk for this voxel is equal to the number of voxels in the neighbourhood that are part of the ROI mask:

δ δ δ X X X Wk = [m 6= 0 and k + m in ROI]

mz =−δ my =−δ mx=−δ

Under our deﬁnition, neighbourhood grey tone diﬀerence si for discretised grey level i can be directly expressed using neighbourhood size Wk of voxel k:

N Xv si = |i − Xk| [Xd(k) = i and Wk 6= 0] k

Consequentially, ni is the total number of voxels with grey level i which have a non-zero neigh- CHAPTER 3. IMAGE FEATURES 117

ni pi si 1 2 2 3 1 0 0.00 0.000 1 2 3 3 2 2 0.50 1.000 i 4 2 4 1 3 1 0.25 0.625 4 1 2 3 4 1 0.25 1.875

(a) Grey levels (b) Neighbourhood grey tone difference matrix Table 3.156 — Original image with grey levels (a) and corresponding neighbourhood grey tone difference matrix (NGTDM) (b). The Nv,c pixels with valid neighbours at distance 1 are located within the rectangle in (a). The grey level voxel count ni, the grey level probability pi = ni/Nv,c, and the neighbourhood grey level difference si for pixels with grey level i are included in the NGTDM. Note that our actual definition deviates from the original definition of Amadasun and King 5 , which is used here. In our definition complete neighbourhood are no longer required. In our definition the NGTDM would be calculated on the entire pixel area, and not solely on those pixels within the rectangle of panel (a). bourhood size.

Many NGTDM-based features depend on the Ng grey level probabilities pi = ni/Nv,c, where Ng P is the number of discretised grey levels in the ROI intensity mask and Nv,c = ni is total number of voxels that have at least one neighbour. If all voxels have at least one neighbour Nv,c = Nv. Furthermore, let Ng,p ≤ Ng be the number of discretised grey levels with pi > 0. In the above example, Ng = 4 and Ng,p = 3.

Aggregating features Three methods can be used to aggregate NGTDMs and arrive at a single feature value. A schematic example was previously shown in Figure 3.4. A feature may be aggregated as follows:

1. Features are computed from 2D matrices and averaged over slices (8QNN).

2. The feature is computed from a single matrix after merging all 2D matrices (62GR).

3. The feature is computed from a 3D matrix (KOBO).

Method 2 involves merging NGTDMs by summing the neighbourhood grey tone diﬀerence si and the number of voxels with a valid neighbourhood ni and grey level i for NGTDMs of the diﬀerent slices. Note that when NGTDMs are merged, Nv,c and pi should be updated based on the merged NGTDM. Feature values may dependent strongly on the aggregation method.

Distances and distance weighting The default neighbourhood is deﬁned using the Chebyshev norm. Manhattan or Euclidean norms may be used as well. This requires a more general deﬁnition for the average grey level Xk:

δ δ δ 1 X X X Xk = Xd(k + m)[kmk ≤ δ and m 6= 0 and k + m in ROI] Wk mz =−δ my =−δ mx=−δ

The neighbourhood size Wk is:

δ δ δ X X X Wk = [kmk ≤ δ and m 6= 0 and k + m in ROI]

mz =−δ my =−δ mx=−δ CHAPTER 3. IMAGE FEATURES 118

As before, [...] is an Iverson bracket. Distance weighting for NGTDM is relatively straightforward. Let w be a weight dependent on m, e.g. w = kmk−1 or w = exp(−kmk2). The average grey level is then:

δ δ δ 1 X X X Xk = w(m)Xd(k + m)[kmk ≤ δ and m 6= 0 and k + m in ROI] Wk mz =−δ my =−δ mx=−δ

The neighbourhood size Wk becomes a general weight:

δ δ δ X X X Wk = w(m)[kmk ≤ δ and m 6= 0 and k + m in ROI]

mz =−δ my =−δ mx=−δ

Employing diﬀerent distance norms and distance weighting is considered non-standard use, and we caution against them due to potential reproducibility issues.

3.10.1 Coarseness QCDE

Grey level differences in coarse textures are generally small due to large-scale patterns. Summing differences gives an indication of the level of the spatial rate of change in intensity 5. Coarseness is defined as: 1 Fngt.coarseness = PNg i=1 pi si

PNg Because i=1 pi si potentially evaluates to 0, the maximum coarseness value is set to an arbitrary number of 106. Amadasun and King originally circumvented this issue by adding a unspeciﬁed small number to the denominator, but an explicit, though arbitrary, maximum value should allow for more consistency.

data aggr. method value tol. consensus

dig. phantom 2D 0.121 — strong dig. phantom 2.5D 0.0285 — strong dig. phantom 3D 0.0296 — very strong config. A 2D 0.00629 0.00046 strong config. A 2.5D 9.06 × 10−5 3.3 × 10−6 strong config. B 2D 0.0168 0.0005 strong config. B 2.5D 0.000314 4 × 10−6 strong config. C 3D 0.000216 4 × 10−6 strong config. D 3D 0.000208 4 × 10−6 strong config. E 3D 0.000188 4 × 10−6 strong

Table 3.157 — Reference values for the coarseness feature. CHAPTER 3. IMAGE FEATURES 119

3.10.2 Contrast 65HE

Contrast depends on the dynamic range of the grey levels as well as the spatial frequency of intensity changes 5. Thus, contrast is deﬁned as:

 Ng Ng   Ng  1 X X 2 1 X Fngt.contrast =  pi1 pi2 (i1 − i2)   si Ng,p (Ng,p − 1) Nv,c i1=1 i2=1 i=1

Grey level probabilities pi1 and pi2 are copies of pi with diﬀerent iterators, i.e. pi1 = pi2 for i1 = i2. The ﬁrst term considers the grey level dynamic range, whereas the second term is a measure for intensity changes within the volume. If Ng,p = 1, Fngt.contrast = 0.

data aggr. method value tol. consensus

dig. phantom 2D 0.925 — strong dig. phantom 2.5D 0.601 — strong dig. phantom 3D 0.584 — very strong config. A 2D 0.107 0.002 strong config. A 2.5D 0.0345 0.0009 strong config. B 2D 0.181 0.001 strong config. B 2.5D 0.0506 0.0005 strong config. C 3D 0.0873 0.0019 strong config. D 3D 0.046 0.0005 strong config. E 3D 0.0752 0.0019 moderate

Table 3.158 — Reference values for the contrast feature.

3.10.3 Busyness NQ30

Textures with large changes in grey levels between neighbouring voxels are said to be busy 5. Busyness was deﬁned as:

PNg i=1 pi si Fngt.busyness = , pi1 6= 0 and pi2 6= 0 PNg PNg i p − i p i1=1 i2=1 1 i1 2 i2

As before, pi1 = pi2 for i1 = i2. The original definition was erroneously formulated as the denominator will always evaluate to 0. Therefore we use a slightly different definition 38:

PNg i=1 pi si Fngt.busyness = , pi1 6= 0 and pi2 6= 0 PNg PNg |i p − i p | i1=1 i2=1 1 i1 2 i2

If Ng,p = 1, Fngt.busyness = 0.

data aggr. method value tol. consensus

dig. phantom 2D 2.99 — strong dig. phantom 2.5D 6.8 — strong dig. phantom 3D 6.54 — very strong conﬁg. A 2D 0.489 0.001 strong conﬁg. A 2.5D 8.84 0.01 strong CHAPTER 3. IMAGE FEATURES 120

config. B 2D 0.2 0.005 strong config. B 2.5D 3.45 0.07 strong config. C 3D 1.39 0.01 very strong config. D 3D 5.14 0.14 strong config. E 3D 4.65 0.1 strong

Table 3.159 — Reference values for the busyness feature.

3.10.4 Complexity HDEZ

Complex textures are non-uniform and rapid changes in grey levels are common 5. Texture complexity is deﬁned as:

Ng Ng 1 X X pi1 si1 + pi2 si2 Fntg.complexity = |i1 − i2| , pi1 6= 0 and pi2 6= 0 Nv,c pi1 + pi2 i1=1 i2=1

As before, pi1 = pi2 for i1 = i2, and likewise si1 = si2 for i1 = i2.

data aggr. method value tol. consensus

dig. phantom 2D 10.4 — strong dig. phantom 2.5D 14.1 — strong dig. phantom 3D 13.5 — very strong config. A 2D 438 9 strong config. A 2.5D 580 19 strong config. B 2D 391 7 strong config. B 2.5D 496 5 strong config. C 3D 1.81 × 103 60 very strong config. D 3D 400 5 strong config. E 3D 574 1 moderate

Table 3.160 — Reference values for the complexity feature.

3.10.5 Strength 1X9X

Amadasun and King 5 deﬁned texture strength as:

PNg PNg (p + p )(i − i )2 i1=1 i2=1 i1 i2 1 2 Fngt.strength = , pi 6= 0 and pi 6= 0 PNg 1 2 i=1 si

PNg As before, pi1 = pi2 for i1 = i2. If i=1 si = 0, Fngt.strength = 0.

data aggr. method value tol. consensus

dig. phantom 2D 2.88 — strong dig. phantom 2.5D 0.741 — strong dig. phantom 3D 0.763 — very strong conﬁg. A 2D 3.33 0.08 strong conﬁg. A 2.5D 0.0904 0.0027 strong CHAPTER 3. IMAGE FEATURES 121

config. B 2D 6.02 0.23 strong config. B 2.5D 0.199 0.009 strong config. C 3D 0.651 0.015 strong config. D 3D 0.162 0.008 strong config. E 3D 0.167 0.006 strong

Table 3.161 — Reference values for the strength feature. CHAPTER 3. IMAGE FEATURES 122

3.11 Neighbouring grey level dependence based featuresREK0

Sun and Wee 69 deﬁned the neighbouring grey level dependence matrix (NGLDM) as an alternative to the grey level co-occurrence matrix. The NGLDM aims to capture the coarseness of the overall texture and is rotationally invariant. NGLDM also involves the concept of a neighbourhood around a central voxel. All voxels within Chebyshev distance δ are considered to belong to the neighbourhood of the center voxel. The discretised grey levels of the center voxel k at position k and a neighbouring voxel m at k + m are said to be dependent if |Xd(k) − Xd(k + m)| ≤ α, with α being a non-negative integer coarseness parameter. The number of grey level dependent voxels j within the neighbourhood is then counted as:

δ δ δ X X X jk = 1 + [|Xd(k) − Xd(k + m)| ≤ α and m 6= 0]

mz =−δ my =−δ mx=−δ

Here, [...] is an Iverson bracket, which is 1 if the aforementioned condition is fulfilled, and 0 otherwise. Note that the minimum dependence jk = 1 and not jk = 0. This is done because some feature definitions require a minimum dependence of 1 or are undefined otherwise. One may therefore also simplify the expression for jk by including the center voxel:

δ δ δ X X X jk = [|Xd(k) − Xd(k + m)| ≤ α]

mz =−δ my =−δ mx=−δ

Dependence jk is iteratively determined for each voxel k in the ROI intensity mask. M is then the Ng ×Nn neighbouring grey level dependence matrix, where Ng is the number of discretised grey levels present in the ROI intensity mask and Nn = max(jk) the maximum grey level dependence count found. Element sij of M is then the number of neighbourhoods with a center voxel with discretised grey level i and a neighbouring voxel dependence j. Furthermore, let Nv be the number PNg PNn of voxels in the ROI intensity mask, and Ns = i=1 j=1 sij the number of neighbourhoods. PNn Marginal sums can likewise be deﬁned. Let si. = j=1 be the number of neighbourhoods with PNg discretised grey level i, and let sj. = i=1 sij be the number of neighbourhoods with dependence j, regardless of grey level. A two dimensional example is shown in Table 3.162. The deﬁnition we actually use deviates from the original by Sun and Wee 69 . Because regions of interest are rarely cuboid, omission of neighbourhoods which contain voxels outside the ROI mask may lead to inconsistent results, especially for larger distance δ. Hence the neighbourhoods of all voxels in the within the ROI intensity mask are considered, and consequently Nv = Ns. Neighbourhood voxels located outside the ROI do not add to dependence j:

δ δ δ X X X jk = [|Xd(k) − Xd(k + m)| ≤ α and k + m in ROI]

mz =−δ my =−δ mx=−δ

Note that while α = 0 is a typical choice for the coarseness parameter, diﬀerent α are possible. Likewise, a typical choice for neighbourhood radius δ is Chebyshev distance δ = 1 but larger values are possible as well.

Aggregating features Three methods can be used to aggregate NGLDMs and arrive at a single feature value. A schematic example was previously shown in Figure 3.4. A feature may be aggregated as follows: CHAPTER 3. IMAGE FEATURES 123

dependence k 0 1 2 3 1 2 2 3 1 0 0 0 0 1 2 3 3 2 0 0 1 1 i 4 2 4 1 3 0 0 1 0 4 1 2 3 4 1 0 0 0

(a) Grey levels (b) Neighbouring grey level dependence matrix

Table 3.162 — Original image with grey levels and pixels with a complete neighbourhood√ within the square (a); corresponding neighbouring grey level dependence matrix for distance d = 2 and coarseness parameter a = 0 (b). Element s(i, j) of the NGLDM indicates the number of neighbourhoods with a center pixel with grey level i and neighbouring grey level dependence k within the image. Note that in our deﬁnition a complete neighbourhood is no longer required. Thus every voxel is considered as a center voxel with a neighbourhood, instead of being constrained to the voxels within the square in panel (a).

1. Features are computed from 2D matrices and averaged over slices (8QNN).

2. The feature is computed from a single matrix after merging all 2D matrices (62GR).

3. The feature is computed from a 3D matrix (KOBO).

Method 2 involves merging NGLDMs by summing the dependence count sij by element over the NGLDM of the diﬀerent slices. Note that when NGLDMs are merged, Nv and Ns should likewise be summed to retain consistency. Feature values may dependent strongly on the aggregation method.

Distances and distance weighting Default neighbourhoods are constructed using the Chebyshev norm, but other norms can be used as well. For this purpose it is useful to generalise the dependence count equation to:

δ δ δ X X X jk = [kmk ≤ δ and |Xd(k) − Xd(k + m)| ≤ α and k + m in ROI]

mz =−δ my =−δ mx=−δ with m the vector between voxels k and m and kmk its length according to the particular norm. In addition, dependence may be weighted by distance. Let w be a weight dependent on m, e.g. w = kmk−1 or w = exp(−kmk2). The dependence of voxel k is then:

δ δ δ X X X jk = w(m)[kmk ≤ δ and |Xd(k) − Xd(k + m)| ≤ α and k + m in ROI]

mz =−δ my =−δ mx=−δ

Employing diﬀerent distance norms and distance weighting is considered non-standard use, and we caution against them due to potential reproducibility issues.

Note on feature references The NGLDM is structured similarly to the GLRLM, GLSZM and GLDZM. NGLDM feature definitions are therefore based on the definitions of GLRLM features, and references may be found in Section 3.7, except for the features originally defined by Sun and Wee 69 . CHAPTER 3. IMAGE FEATURES 124

3.11.1 Low dependence emphasis SODN

This feature emphasises low neighbouring grey level dependence counts. Sun and Wee 69 refer to this feature as small number emphasis. It is deﬁned as:

Nn 1 X s.j F = ngl.lde N j2 s j=1

data aggr. method value tol. consensus

dig. phantom 2D 0.158 — strong dig. phantom 2.5D 0.159 — strong dig. phantom 3D 0.045 — very strong config. A 2D 0.281 0.003 strong config. A 2.5D 0.243 0.004 strong config. B 2D 0.31 0.001 strong config. B 2.5D 0.254 0.002 strong config. C 3D 0.137 0.003 very strong config. D 3D 0.0912 0.0007 strong config. E 3D 0.118 0.001 strong

Table 3.163 — Reference values for the low dependence emphasis feature.

3.11.2 High dependence emphasis IMOQ

This feature emphasises high neighbouring grey level dependence counts. Sun and Wee 69 refer to this feature as large number emphasis. It is deﬁned as:

N 1 Xn F = j2s ngl.hde N .j s j=1

data aggr. method value tol. consensus

dig. phantom 2D 19.2 — strong dig. phantom 2.5D 18.8 — strong dig. phantom 3D 109 — very strong config. A 2D 14.8 0.1 strong config. A 2.5D 16.1 0.2 strong config. B 2D 17.3 0.2 strong config. B 2.5D 19.6 0.2 strong config. C 3D 126 2 strong config. D 3D 223 5 strong config. E 3D 134 3 strong

Table 3.164 — Reference values for the high dependence emphasis feature. CHAPTER 3. IMAGE FEATURES 125

3.11.3 Low grey level count emphasis TL9H

This feature is a grey level analogue to low dependence emphasis. Instead of low neighbouring grey level dependence counts, low grey levels are emphasised. The feature is deﬁned as:

Ng 1 X si. F = ngl.lgce N i2 s i=1

data aggr. method value tol. consensus

dig. phantom 2D 0.702 — strong dig. phantom 2.5D 0.693 — strong dig. phantom 3D 0.693 — very strong config. A 2D 0.0233 0.0003 strong config. A 2.5D 0.0115 0.0003 strong config. B 2D 0.0286 0.0004 strong config. B 2.5D 0.0139 0.0005 strong config. C 3D 0.0013 4 × 10−5 strong config. D 3D 0.0168 0.0009 strong config. E 3D 0.0154 0.0007 strong

Table 3.165 — Reference values for the low grey level count emphasis feature.

3.11.4 High grey level count emphasis OAE7

The high grey level count emphasis feature is a grey level analogue to high dependence emphasis. The feature emphasises high grey levels, and is deﬁned as:

Ng 1 X F = i2s ngl.hgce N i. s i=1

data aggr. method value tol. consensus

dig. phantom 2D 7.49 — strong dig. phantom 2.5D 7.66 — strong dig. phantom 3D 7.66 — very strong config. A 2D 446 2 strong config. A 2.5D 466 2 strong config. B 2D 359 10 strong config. B 2.5D 375 11 strong config. C 3D 1.57 × 103 10 strong config. D 3D 364 16 strong config. E 3D 502 8 strong

Table 3.166 — Reference values for the high grey level count emphasis feature. CHAPTER 3. IMAGE FEATURES 126

3.11.5 Low dependence low grey level emphasis EQ3F

This feature emphasises neighbouring grey level dependence counts in the upper left quadrant of the NGLDM, where low dependence counts and low grey levels are located. It is deﬁned as:

Ng Nn 1 X X sij F = ngl.ldlge N i2j2 s i=1 j=1

data aggr. method value tol. consensus

dig. phantom 2D 0.0473 — strong dig. phantom 2.5D 0.0477 — strong dig. phantom 3D 0.00963 — very strong config. A 2D 0.0137 0.0002 strong config. A 2.5D 0.00664 0.0002 strong config. B 2D 0.0203 0.0003 strong config. B 2.5D 0.00929 0.00026 strong config. C 3D 0.000306 1.2 × 10−5 strong config. D 3D 0.00357 4 × 10−5 strong config. E 3D 0.00388 4 × 10−5 strong

Table 3.167 — Reference values for the low dependence low grey level emphasis feature.

3.11.6 Low dependence high grey level emphasis JA6D

This feature emphasises neighbouring grey level dependence counts in the lower left quadrant of the NGLDM, where low dependence counts and high grey levels are located. The feature is deﬁned as:

Ng Nn 2 1 X X i sij F = ngl.ldhge N j2 s i=1 j=1

data aggr. method value tol. consensus

dig. phantom 2D 3.06 — strong dig. phantom 2.5D 3.07 — strong dig. phantom 3D 0.736 — very strong config. A 2D 94.2 0.4 strong config. A 2.5D 91.9 0.5 strong config. B 2D 78.9 2.2 strong config. B 2.5D 73.4 2.1 strong config. C 3D 141 2 strong config. D 3D 18.9 1.1 strong config. E 3D 36.7 0.5 strong

Table 3.168 — Reference values for the low dependence high grey level emphasis feature. CHAPTER 3. IMAGE FEATURES 127

3.11.7 High dependence low grey level emphasis NBZI

This feature emphasises neighbouring grey level dependence counts in the upper right quadrant of the NGLDM, where high dependence counts and low grey levels are located. The feature is deﬁned as:

Ng Nn 2 1 X X j sij F = ngl.hdlge N i2 s i=1 j=1

data aggr. method value tol. consensus

dig. phantom 2D 17.6 — strong dig. phantom 2.5D 17.2 — strong dig. phantom 3D 102 — very strong config. A 2D 0.116 0.001 strong config. A 2.5D 0.0674 0.0004 strong config. B 2D 0.108 0.003 strong config. B 2.5D 0.077 0.0019 strong config. C 3D 0.0828 0.0003 strong config. D 3D 0.798 0.072 strong config. E 3D 0.457 0.031 strong

Table 3.169 — Reference values for the high dependence low grey level emphasis feature.

3.11.8 High dependence high grey level emphasis 9QMG

The high dependence high grey level emphasis feature emphasises neighbouring grey level dependence counts in the lower right quadrant of the NGLDM, where high dependence counts and high grey levels are located. The feature is deﬁned as:

Ng N 1 X Xn F = i2j2s ngl.hdhge N ij s i=1 j=1

data aggr. method value tol. consensus

dig. phantom 2D 49.5 — strong dig. phantom 2.5D 50.8 — strong dig. phantom 3D 235 — very strong config. A 2D 7.54 × 103 60 strong config. A 2.5D 8.1 × 103 60 strong config. B 2D 7.21 × 103 130 strong config. B 2.5D 7.97 × 103 150 strong config. C 3D 2.27 × 105 3 × 103 strong config. D 3D 9.28 × 104 1.3 × 103 strong config. E 3D 7.6 × 104 600 strong

Table 3.170 — Reference values for the high dependence high grey level emphasis feature. CHAPTER 3. IMAGE FEATURES 128

3.11.9 Grey level non-uniformity FP8K

Grey level non-uniformity assesses the distribution of neighbouring grey level dependence counts over the grey values. The feature value is low when dependence counts are equally distributed along grey levels. The feature is deﬁned as:

Ng 1 X F = s2 ngl.glnu N i. s i=1

data aggr. method value tol. consensus

dig. phantom 2D 10.2 — strong dig. phantom 2.5D 37.9 — strong dig. phantom 3D 37.9 — very strong config. A 2D 757 1 strong config. A 2.5D 1.72 × 104 100 strong config. B 2D 216 3 strong config. B 2.5D 4.76 × 103 50 strong config. C 3D 6.42 × 103 10 strong config. D 3D 1.02 × 104 300 strong config. E 3D 8.17 × 103 130 strong

Table 3.171 — Reference values for the grey level non-uniformity feature.

3.11.10 Normalised grey level non-uniformity 5SPA

This is a normalised version of the grey level non-uniformity feature. It is deﬁned as:

Ng 1 X F = s2 ngl.glnu.norm N 2 i. s i=1

The normalised grey level non-uniformity computed from a single 3D NGLDM matrix is equivalent to the intensity histogram uniformity feature 81.

data aggr. method value tol. consensus

dig. phantom 2D 0.562 — strong dig. phantom 2.5D 0.512 — strong dig. phantom 3D 0.512 — very strong config. A 2D 0.151 0.003 strong config. A 2.5D 0.15 0.002 strong config. B 2D 0.184 0.001 strong config. B 2.5D 0.174 0.001 strong config. C 3D 0.14 0.003 very strong config. D 3D 0.229 0.003 strong config. E 3D 0.184 0.001 strong

Table 3.172 — Reference values for the normalised grey level non-uniformity feature. CHAPTER 3. IMAGE FEATURES 129

3.11.11 Dependence count non-uniformity Z87G

This features assesses the distribution of neighbouring grey level dependence counts over the diﬀer- ent dependence counts. The feature value is low when dependence counts are equally distributed. Sun and Wee 69 refer to this feature as number non-uniformity. It is deﬁned as:

N 1 Xn F = s2 ngl.dcnu N .j s j=1

-5mm

data aggr. method value tol. consensus

dig. phantom 2D 3.96 — strong dig. phantom 2.5D 12.4 — strong dig. phantom 3D 4.86 — very strong config. A 2D 709 2 strong config. A 2.5D 1.75 × 104 100 strong config. B 2D 157 1 strong config. B 2.5D 3.71 × 103 30 strong config. C 3D 2.45 × 103 60 strong config. D 3D 1.84 × 103 30 strong config. E 3D 2.25 × 103 30 strong

Table 3.173 — Reference values for the dependence count non-uniformity feature.

3.11.12 Normalised dependence count non-uniformity OKJI

This is a normalised version of the dependence count non-uniformity feature. It is deﬁned as:

N 1 Xn F = s2 ngl.dcnu.norm N 2 .j s i=1

data aggr. method value tol. consensus

dig. phantom 2D 0.212 — strong dig. phantom 2.5D 0.167 — strong dig. phantom 3D 0.0657 — very strong config. A 2D 0.175 0.001 strong config. A 2.5D 0.153 0.001 strong config. B 2D 0.179 0.001 strong config. B 2.5D 0.136 0.001 strong config. C 3D 0.0532 0.0005 strong config. D 3D 0.0413 0.0003 strong config. E 3D 0.0505 0.0003 strong

Table 3.174 — Reference values for the normalised dependence count non-uniformity feature. CHAPTER 3. IMAGE FEATURES 130

3.11.13 Dependence count percentage 6XV8

This feature measures the fraction of the number of realised neighbourhoods and the maximum number of potential neighbourhoods. Dependence count percentage may be completely omitted as it evaluates to 1 when complete neighbourhoods are not required, as is the case under our deﬁnition. It is deﬁned as:

Ns Fngl.dc.perc = Nv

data aggr. method value tol. consensus

dig. phantom 2D 1 — strong dig. phantom 2.5D 1 — moderate dig. phantom 3D 1 — strong config. A 2D 1 — moderate config. A 2.5D 1 — strong config. B 2D 1 — moderate config. B 2.5D 1 — moderate config. C 3D 1 — strong config. D 3D 1 — strong config. E 3D 1 — moderate

Table 3.175 — Reference values for the dependence count percentage feature.

3.11.14 Grey level variance 1PFV

This feature estimates the variance in dependence counts over the grey levels. Let pij = sij/Ns be the joint probability estimate for ﬁnding discretised grey level i with dependence j. The feature is then deﬁned as:

Ng Nn X X 2 Fngl.gl.var = (i − µ) pij i=1 j=1

PNg PNn Here, µ = i=1 j=1 i pij.

data aggr. method value tol. consensus

dig. phantom 2D 2.7 — strong dig. phantom 2.5D 3.05 — strong dig. phantom 3D 3.05 — very strong config. A 2D 31.1 0.5 strong config. A 2.5D 22.8 0.6 strong config. B 2D 25.3 0.4 strong config. B 2.5D 18.7 0.2 strong config. C 3D 81.1 2.1 very strong config. D 3D 21.7 0.4 strong config. E 3D 30.4 0.8 strong CHAPTER 3. IMAGE FEATURES 131

Table 3.176 — Reference values for the grey level variance feature.

5mm

3.11.15 Dependence count variance DNX2

This feature estimates the variance in dependence counts over the diﬀerent possible dependence counts. As before let pij = sij/Ns. The feature is deﬁned as:

Ng Nn X X 2 Fngl.dc.var = (j − µ) pij i=1 j=1

PNg PNn Mean dependence count is deﬁned as µ = i=1 j=1 j pij.

data aggr. method value tol. consensus

dig. phantom 2D 2.73 — strong dig. phantom 2.5D 3.27 — strong dig. phantom 3D 22.1 — very strong config. A 2D 3.12 0.02 strong config. A 2.5D 3.37 0.01 strong config. B 2D 4.02 0.05 strong config. B 2.5D 4.63 0.06 strong config. C 3D 39.2 0.1 strong config. D 3D 63.9 1.3 strong config. E 3D 39.4 1 strong

Table 3.177 — Reference values for the dependence count variance feature.

3.11.16 Dependence count entropy FCBV

69 This feature is referred to as entropy by Sun and Wee . Let pij = sij/Ns. Dependence count entropy is then deﬁned as:

Ng N X Xn Fngl.dc.entr = − pij log2 pij i=1 j=1

This deﬁnition remedies an error in the deﬁnition of Sun and Wee 69 , where the term within the logarithm is dependence count sij instead of count probability pij.

data aggr. method value tol. consensus

dig. phantom 2D 2.71 — strong dig. phantom 2.5D 3.36 — strong dig. phantom 3D 4.4 — very strong config. A 2D 5.76 0.02 strong config. A 2.5D 5.93 0.02 strong config. B 2D 5.38 0.01 strong CHAPTER 3. IMAGE FEATURES 132

config. B 2.5D 5.78 0.01 strong config. C 3D 7.54 0.03 very strong config. D 3D 6.98 0.01 strong config. E 3D 7.06 0.02 strong

Table 3.178 — Reference values for the dependence count entropy feature.

3.11.17 Dependence count energy CAS9

69 This feature is called second moment by Sun and Wee . Let pij = sij/Ns. Then dependence count energy is deﬁned as:

Ng Nn X X 2 Fngl.dc.energy = pij i=1 j=1

This definition also remedies an error in the original definition, where squared dependence count 2 sij is divided by Ns only, thus leaving a major volume dependency. In the definition given here, 2 2 sij is normalised by Ns through the use of count probability pij.

data aggr. method value tol. consensus

dig. phantom 2D 0.17 — strong dig. phantom 2.5D 0.122 — strong dig. phantom 3D 0.0533 — very strong config. A 2D 0.0268 0.0004 strong config. A 2.5D 0.0245 0.0003 strong config. B 2D 0.0321 0.0002 strong config. B 2.5D 0.0253 0.0001 moderate config. C 3D 0.00789 0.00011 strong config. D 3D 0.0113 0.0002 strong config. E 3D 0.0106 0.0001 strong

Table 3.179 — Reference values for the dependence count energy feature. Chapter 4

Radiomics reporting guidelines and nomenclature

Reliable and complete reporting is necessary to ensure reproducibility and validation of results. To help provide a complete report on image processing and image biomarker extraction, we present the guidelines below, as well as a nomenclature system to uniquely features.

4.1 Reporting guidelines

These guidelines are partially based on the work of Lambin et al. 41 , Sanduleanu et al. 57 , Sollini et al. 64 , Traverso et al. 73 . Additionally, guidelines are derived from the image processing and feature calculation steps described within this document. An earlier version was reported elsewhere 79.

topic item description

Patient

Region of interest1 1 Describe the region of interest that is being im- aged. Patient preparation 2a Describe speciﬁc instructions given to patients prior to image acquisition, e.g. fasting prior to imaging. 2b Describe administration of drugs to the patient prior to image acquisition, e.g. muscle relaxants. 2c Describe the use of speciﬁc equipment for patient comfort during scanning, e.g. ear plugs. Radioactive tracer PET, SPECT 3a Describe which radioactive tracer was administered to the patient, e.g. 18F-FDG. PET, SPECT 3b Describe the administration method. PET, SPECT 3c Describe the injected activity of the radioactive tracer at administration. PET, SPECT 3d Describe the uptake time prior to image acquisition. continued on next page

1Also referred to as volume of interest.

133 CHAPTER 4. RADIOMICS REPORTING GUIDELINES AND NOMENCLATURE 134

topic item description

PET, SPECT 3e Describe how competing substance levels were controlled.2 Contrast agent 4a Describe which contrast agent was administered to the patient. 4b Describe the administration method. 4c Describe the injected quantity of contrast agent. 4d Describe the uptake time prior to image acquisition. 4e Describe how competing substance levels were controlled. Comorbidities 5 Describe if the patients have comorbidities that aﬀect imaging.3 Acquisition4

Acquisition protocol 6 Describe whether a standard imaging protocol was used, and where its description may be found. Scanner type 7 Describe the scanner type(s) and vendor(s) used in the study. Imaging modality 8 Clearly state the imaging modality that was used in the study, e.g. CT, MRI. Static/dynamic scans 9a State if the scans were static or dynamic. Dynamic scans 9b Describe the acquisition time per time frame. Dynamic scans 9c Describe any temporal modelling technique that was used. Scanner calibration 10 Describe how and when the scanner was calibrated. Patient instructions 11 Describe specific instructions given to the patient during acquisition, e.g. breath holding. Anatomical motion cor- 12 Describe the method used to minimise the effect rection of anatomical motion. Scan duration 13 Describe the duration of the complete scan or the time per bed position. Tube voltage CT 14 Describe the peak kilo voltage output of the X-ray source. Tube current CT 15 Describe the tube current in mA. Time-of-flight PET 16 State if scanner time-of-flight capabilities are used during acquisition. RF coil MRI 17 Describe what kind RF coil used for acquisition, incl. vendor. Scanning sequence MRI 18a Describe which scanning sequence was acquired. MRI 18b Describe which sequence variant was acquired. MRI 18c Describe which scan options apply to the current sequence, e.g. flow compensation, cardiac gating. continued on next page

2An example is glucose present in the blood which competes with the uptake of 18F-FDG tracer in tumour tissue. To reduce competition with the tracer, patients are usually asked to fast for several hours and a blood glucose measurement may be conducted prior to tracer administration. 3An example of a comorbidity that may aﬀect image quality in 18F-FDG PET scans are type I and type II diabetes melitus, as well as kidney failure. 4Many acquisition parameters may be extracted from DICOM header meta-data, or calculated from them. CHAPTER 4. RADIOMICS REPORTING GUIDELINES AND NOMENCLATURE 135

topic item description

Repetition time MRI 19 Describe the time in ms between subsequent pulse sequences. Echo time MRI 20 Describe the echo time in ms. Echo train length MRI 21 Describe the number of lines in k-space that are acquired per excitation pulse. Inversion time MRI 22 Describe the time in ms between the middle of the inverting RF pulse to the middle of the excitation pulse. Flip angle MRI 23 Describe the flip angle produced by the RF pulses. Acquisition type MRI 24 Describe the acquisition type of the MRI scan, e.g. 3D. k-space traversal MRI 25 Describe the acquisition trajectory of the k-space. Number of averages/ ex- MRI 26 Describe the number of times each point in k- citations space is sampled. Magnetic field strength MRI 27 Describe the nominal strength of the MR magnetic field. Reconstruction5

In-plane resolution 28 Describe the distance between pixels, or alternatively the ﬁeld of view and matrix size. Image slice thickness 29 Describe the slice thickness. Image slice spacing 30 Describe the distance between image slices.6 Convolution kernel CT 31a Describe the convolution kernel used to recon- struct the image. CT 31b Describe settings pertaining to iterative reconstruction algorithms. Exposure CT 31c Describe the exposure (in mAs) in slices containing the region of interest. Reconstruction method PET 32a Describe which reconstruction method was used, e.g. 3D OSEM. PET 32b Describe the number of iterations for iterative reconstruction. PET 32c Describe the number of subsets for iterative reconstruction. Point spread function PET 33 Describe if and how point-spread function model- modelling ling was performed. Image corrections PET 34a Describe if and how attenuation correction was performed. PET 34b Describe if and how other forms of correction were performed, e.g. scatter correction, randoms correction, dead time correction etc. Reconstruction method MRI 35a Describe the reconstruction method used to recon- struct the image from the k-space information. MRI 35b Describe any artifact suppression methods used during reconstruction to suppress artifacts due to undersampling of k-space. continued on next page

5Many reconstruction parameters may be extracted from DICOM header meta-data. 6Spacing between image slicing is commonly, but not necessarily, the same as the slice thickness,. CHAPTER 4. RADIOMICS REPORTING GUIDELINES AND NOMENCLATURE 136

topic item description

Diﬀusion-weighted ima- DWI-MRI 36 Describe the b-values used for diﬀusion-weighting. ging Image registration

Registration method 37 Describe the method used to register multi- modality imaging. Image processing - data conversion

SUV normalisation PET 38 Describe which standardised uptake value (SUV) normalisation method is used. ADC computation DWI-MRI 39 Describe how apparent diﬀusion coeﬃcient (ADC) values were calculated. Other data conversions 40 Describe any other conversions that are performed to generate e.g. perfusion maps. Image processing - post-acquisition processing

Anti-aliasing 41 Describe the method used to deal with anti- aliasing when down-sampling during interpolation. Noise suppression 42 Describe methods used to suppress image noise. Post-reconstruction PET 43 Describe the width of the Gaussian ﬁlter (FWHM) smoothing ﬁlter to spatially smooth intensities. Skull stripping MRI (brain) 44 Describe method used to perform skull stripping. Non-uniformity correc- MRI 45 Describe the method and settings used to perform tion7 non-uniformity correction. Intensity normalisation 46 Describe the method and settings used to normal- ise intensity distributions within a patient or patient cohort. Other post-acquisition 47 Describe any other methods that were used to pro- processing methods cess the image and are not mentioned separately in this list. Segmentation

Segmentation method 48a Describe how regions of interest were segmented, e.g. manually. 48b Describe the number of experts, their expertise and consensus strategies for manual delineation. 48c Describe methods and settings used for semi- automatic and fully automatic segmentation. 48d Describe which image was used to deﬁne segmentation in case of multi-modality imaging. Conversion to mask 49 Describe the method used to convert polygonal or mesh-based segmentations to a voxel-based mask. Image processing - image interpolation

Interpolation method 50a Describe which interpolation algorithm was used to interpolate the image. 50b Describe how the position of the interpolation grid was deﬁned, e.g. align by center. continued on next page

7Also known as bias-ﬁeld correction. CHAPTER 4. RADIOMICS REPORTING GUIDELINES AND NOMENCLATURE 137

topic item description

50c Describe how the dimensions of the interpolation grid were deﬁned, e.g. rounded to nearest integer. 50d Describe how extrapolation beyond the original image was handled. Voxel dimensions 51 Describe the size of the interpolated voxels. Intensity rounding CT 52 Describe how fractional Hounsﬁeld Units are rounded to integer values after interpolation. Image processing - ROI interpolation

Interpolation method 53 Describe which interpolation algorithm was used to interpolate the region of interest mask. Partially masked voxels 54 Describe how partially masked voxels after interpolation are handled. Image processing - re-segmentation

Re-segmentation meth- 55 Describe which methods and settings are used to ods re-segment the ROI intensity mask. Image processing - discretisation

Discretisation method8 56a Describe the method used to discretise image intensities. 56b Describe the number of bins (FBN) or the bin size (FBS) used for discretisation. 56c Describe the lowest intensity in the ﬁrst bin for FBS discretisation.9 Image processing - image transformation

Image ﬁlter10 57 Describe the methods and settings used to ﬁlter images, e.g. Laplacian-of-Gaussian. Image biomarker computation

Biomarker set 58 Describe which set of image biomarkers is computed and refer to their deﬁnitions or provide these. IBSI compliance 59 State if the software used to extract the set of image biomarkers is able to reproduce the IBSI feature reference values.11 Robustness 60 Describe how robustness of the image biomarkers was assessed, e.g. test-retest analysis. Software availability 61 Describe which software and version was used to compute image biomarkers. Image biomarker computation - texture parameters

Texture matrix aggrega- 62 Deﬁne how texture-matrix based biomarkers were tion computed from underlying texture matrices. continued on next page

8Discretisation may be performed separately to create intensity-volume histograms. If this is indeed the case, this should be described as well. 9This is typically set by range re-segmentation. 10The IBSI has not introduced image transformation into the standardised image processing scheme, and is in the process of benchmarking various common ﬁlters. This section may therefore be expanded in the future. 11A software is compliant if and only if it is able to reproduce image biomarker reference values for the digital phantom and for one or more image processing conﬁgurations using the radiomics CT phantom. Reviewers may demand that you provide the IBSI compliance spreadsheet for your software. CHAPTER 4. RADIOMICS REPORTING GUIDELINES AND NOMENCLATURE 138

topic item description

Distance weighting 63 Define how CM, RLM, NGTDM and NGLDM weight distances, e.g. no weighting. CM symmetry 64 Define whether symmetric or asymmetric co- occurrence matrices were computed. CM distance 65 Define the (Chebyshev) distance at which co- occurrence of intensities is determined, e.g. 1. SZM linkage distance 66 Define the distance and distance norm for which voxels with the same intensity are considered to belong to the same zone for the purpose of constructing an SZM, e.g. Chebyshev distance of 1. DZM linkage distance 67 Define the distance and distance norm for which voxels with the same intensity are considered to belong to the same zone for the purpose of constructing a DZM, e.g. Chebyshev distance of 1. DZM zone distance norm 68 Define the distance norm for determining the distance of zones to the border of the ROI, e.g. Man- hattan distance. NGTDM distance 69 Define the neighbourhood distance and distance norm for the NGTDM, e.g. Chebyshev distance of 1. NGLDM distance 70 Define the neighbourhood distance and distance norm for the NGLDM, e.g. Chebyshev distance of 1. NGLDM coarseness 71 Define the coarseness parameter for the NGLDM, e.g. 0. Machine learning and radiomics analysis

Diagnostic and pro- 72 See the TRIPOD guidelines for reporting on diagnostic modelling gnostic and prognostic modelling. Comparison with known 73 Describe where performance of radiomics models factors is compared with known (clinical) factors. Multicollinearity 74 Describe where the multicollinearity between image biomarkers in the signature is assessed. Model availability 75 Describe where radiomics models with the necessary pre-processing information may be found. Data availability 76 Describe where imaging data and relevant meta- data used in the study may be found.

Table 4.1 — Guidelines for reporting on radiomic studies. Not all items may be applicable.

4.2 Feature nomenclature

Image features may be extracted using a variety of different settings, and may even share the same name. A feature nomenclature is thus required. Let us take the example of differentiating the following features: i) intensity histogram-based entropy, discretised using a fixed bin size algorithm with 25 HU bins, extracted from a CT image; and ii) grey level run length matrix entropy, discretised using a fixed bin number algorithm with 32 bins, extracted from a PET image. To refer to both as entropy would be ambiguous, whereas to add a full textual description would be cumber- some. In the nomenclature proposed below, the features would be called entropyIH, CT, FBS:25HU CHAPTER 4. RADIOMICS REPORTING GUIDELINES AND NOMENCLATURE 139

and entropyRLM, PET, FBN:32, respectively. Features are thus indicated by a feature name and a subscript. As the nomenclature is designed to both concise and complete, only details for which ambiguity may exist are to be explicitly incorporated in the subscript. The subscript of a feature name may contain the following items to address ambiguous naming:

1. An abbreviation of the feature family (required).

2. The aggregation method of a feature (optional).

3. A descriptor describing the modality the feature is based on, the specific channel (for microscopy images), the specific imaging data (in the case of repeat imaging or delta-features) sets, conversions (such as SUV and SUL), and/or the specific ROI. For example, one could write PET:SUV to separate it from CT and PET:SUL features (optional).

4. Spatial ﬁlters and settings (optional).

5. The interpolation algorithm and uniform interpolation grid spacing (optional).

6. The re-segmentation range and outlier ﬁltering (optional).

7. The discretisation method and relevant discretisation parameters, i.e. number of bins or bin size (optional).

8. Feature speciﬁc parameters, such as distance for some texture features (optional).

Optional descriptors are only added to the subscript if there are multiple possibilities. For example, if only CT data is used, adding the modality to the subscript is not required. Nonetheless, such details must be reported as well (see section 4.1). The sections below have tables with permanent IBSI identiﬁers for concepts that were deﬁned within this document.

4.2.1 Abbreviating feature families

The following is a list of the feature families in this document and their suggested abbreviations:

feature family abbreviation morphology MORPH HCUG local intensity LI 9ST6 intensity-based statistics IS, STAT UHIW intensity histogram IH ZVCW intensity-volume histogram IVH P88C grey level co-occurrence matrix GLCM, CM LFYI grey level run length matrix GLRLM, RLM TP0I grey level size zone matrix GLSZM, SZM 9SAK grey level distance zone matrix GLDZM, DZM VMDZ neighbourhood grey tone diﬀerence matrix NGTDM IPET neighbouring grey level dependence matrix NGLDM REK0

4.2.2 Abbreviating feature aggregation

The following is a list of feature families and the possible aggregation methods: CHAPTER 4. RADIOMICS REPORTING GUIDELINES AND NOMENCLATURE 140

morphology, LI – features are 3D by deﬁnition DHQ4

IS, IH, IVH 2D averaged over slices (rare) 3IDG –, 3D calculated over the volume (default) DHQ4

GLCM, GLRLM 2D:avg averaged over slices and directions BTW3 2D:mrg, 2D:smrg merged directions per slice and averaged SUJT 2.5D:avg, 2.5D:dmrg merged per direction and averaged JJUI 2.5D:mrg, 2.5D:vmrg merged over all slices ZW7Z 3D:avg averaged over 3D directions ITBB 3D:mrg merged 3D directions IAZD

GLSZM, GLDZM, NGTDM, NGLDM 2D averaged over slices 8QNN 2.5D merged over all slices 62GR 3D calculated from single 3D matrix KOBO

In the list above, ’–’ signiﬁes an empty entry which does not need to be added to the subscript. The following examples highlight the nomenclature used above:

• joint maximumCM, 2D:avg: GLCM-based joint maximum feature, calculated by averaging the feature for every in-slice GLCM.

• short runs emphasisRLM, 3D:mrg: RLM-based short runs emphasis feature, calculated from an RLM that was aggregated by merging the RLM of each 3D direction.

• meanIS: intensity statistical mean feature, calculated over the 3D ROI volume.

• grey level varianceSZM, 2D: SZM-based grey level variance feature, calculated by averaging the feature value from the SZM in each slice over all the slices.

4.2.3 Abbreviating interpolation

The following is a list of interpolation methods and the suggested notation. Note that # is the interpolation spacing, including units, and dim is 2D for interpolation with the slice plane and 3D for volumetric interpolation.

interpolation method notation

none INT:– nearest neighbour interpolation NNB:dim:# linear interpolation LIN:dim:# cubic convolution interpolation CCI:dim:# cubic spline interpolation CSI:dim:#, SI3:dim:#

The dimension attribute and interpolation spacing may be omitted if this is clear from the context. The following examples highlight the nomenclature introduced above:

• meanIS, LIN:2D:2mm: intensity statistical mean feature, calculated after bilinear interpolation with the slice planes to uniform voxel sizes of 2mm. CHAPTER 4. RADIOMICS REPORTING GUIDELINES AND NOMENCLATURE 141

• meanIH, NNB:3D:1mm: intensity histogram mean feature, calculated after trilinear interpolation to uniform voxel sizes of 1mm.

• joint maximumCM, 2D:mrg, CSI:2D:2mm: GLCM-based joint maximum feature, calculated by ﬁrst merging all GLCM within a slice to single GLCM, calculating the feature and then averaging the feature values over the slices. GLCMs were determined in the image interpolated within the slice plane to 2 × 2mm voxels using cubic spline interpolation.

4.2.4 Describing re-segmentation

Re-segmentation can be noted as follows:

re-segmentation method notation

none RS:– range RS:[#,#] USB3 outlier ﬁltering RS:#σ 7ACA

In the table above # signify numbers. A re-segmentation range can be half-open, i.e. RS:[#,∞). Re-segmentation methods may be combined, i.e. both range and outlier ﬁltering methods may be used. This is noted as RS:[#,#]+#σ or RS:#σ+[#,#]. The following are examples of the application of the above notation:

• meanIS, CT, RS:[-200,150]: intensity statistical mean feature, based on an ROI in a CT image that was re-segmented within a [-200,150] HU range.

• meanIS, PET:SUV, RS:[3,∞): intensity statistical mean feature, based on an ROI in a PET image with SUV values, that was re-segmented to contain only SUV of 3 and above.

• meanIS, MRI:T1, RS:3σ: intensity statistical mean feature, based on an ROI in a T1-weighted MR image where the ROI was re-segmented by removing voxels with an intensity outside a µ ± 3σ range.

4.2.5 Abbreviating discretisation

The following is a list of discretisation methods and the suggested notation. Note that # is the value of the relevant discretisation parameter, e.g. number of bins or bin size, including units.

discretisation method notation

none DIS:– ﬁxed bin size FBS:# Q3RU ﬁxed bin number FBN:# K15C histogram equalisation EQ:# Lloyd-Max, minimum mean squared LM:#, MMS:#

In the table above, # signify numbers such as the number of bins or their width. Histogram equalisation of the ROI intensities can be performed before the ”none”, ”fixed bin size”, ”fixed bin number” or ”Lloyd-Max, minimum mean squared” algorithms defined above, with # specifying the number of bins in the histogram to be equalised. The following are examples of the application of the above notation:

• meanIH,PET:SUV,RS[0,∞],FBS:0.2: intensity histogram mean feature, based on an ROI in a SUV-PET image, with bin-width of 0.2 SUV, and binning from 0.0 SUV. CHAPTER 4. RADIOMICS REPORTING GUIDELINES AND NOMENCLATURE 142

• grey level varianceSZM,MR:T1,RS:3σ,FBN:64: size zone matrix-based grey level variance feature, based on an ROI in a T1-weighted MR image, with 3σ re-segmentation and subsequent binning into 64 bins.

4.2.6 Abbreviating feature-speciﬁc parameters

Some features and feature families require additional parameters, which may be varied. These are the following:

grey level co-occurrence matrix

co-occurrence matrix symmetry –, SYM symmetrical co-occurrence matrices ASYM asymmetrical co-occurrence matrices (not recommended)

distance

δ:#, δ-∞:# Chebyshev (`∞) norm with distance # (default) PVMT

δ-2:# Euclidean (`2) norm with distance # G9EV

δ-1:# Manhattan (`1) norm with distance # LIFZ

distance weighting –, w:1 no weighting (default) w:f weighting with function f

grey level run length matrix

distance weighting –, w:1 no weighting (default) w:f weighting with function f

grey level size zone matrix

linkage distance

δ:#, δ-∞:# Chebyshev (`∞) norm with distance (default) # PVMT

δ-2:# Euclidean (`2) norm with distance # G9EV

δ-1:# Manhattan (`1) norm with distance # LIFZ

grey level distance zone matrix

linkage distance

δ:#, δ-∞:# Chebyshev (`∞) norm with distance (default) # PVMT

δ-2:# Euclidean (`2) norm with distance # G9EV

δ-1:# Manhattan (`1) norm with distance # LIFZ

zone distance norm

l-∞:# Chebyshev (`∞) norm PVMT

l-2:# Euclidean (`2) norm G9EV

–, l-1:# Manhattan (`1) norm (default) LIFZ

neighbourhood grey tone diﬀerence matrix

distance

δ:#, δ-∞:# Chebyshev (`∞) norm with distance # (default) PVMT continued on next page CHAPTER 4. RADIOMICS REPORTING GUIDELINES AND NOMENCLATURE 143

δ-2:# Euclidean (`2) norm with distance # G9EV

δ-1:# Manhattan (`1) norm with distance # LIFZ

distance weighting –, w:1 no weighting (default) w:f weighting with function f

neighbouring grey level dependence matrix

dependence coarseness α:# dependence coarseness parameter with value #

distance

δ:#, δ-∞:# Chebyshev (`∞) norm with distance # (default) PVMT

δ-2:# Euclidean (`2) norm with distance # G9EV

δ-1:# Manhattan (`1) norm with distance # LIFZ

distance weighting –, w:1 no weighting (default) w:f weighting with function f

In the above table, # represents numbers. Chapter 5

Reference data sets

Reference values for features were obtained using a digital image phantom and the CT image of a lung cancer patient, which are described below. The same data sets can be used to verify radiomics software implementations. The data sets themselves may be found here: https://github.com/theibsi/data sets.

5.1 Digital phantom

A small digital phantom was developed to derive image features manually and compare these values with values obtained from radiomics software implementations. The phantom is shown in ﬁgure 5.1. The phantom has the following characteristics:

• The phantom consists of 5 × 4 × 4 (x, y, z) voxels.

• A slice consists of the voxels in (x, y) plane for a particular slice at position z. Slices are therefore stacked in the z direction.

• Voxels are 2.0 × 2.0 × 2.0 mm in size.

• Not all voxels are included in the region of interest. Several excluded voxels are located on the outside of the ROI, and one internal voxel was excluded as well. Voxels excluded from the ROI are shown in blue in ﬁgure 5.1.

• Some intensities are not present in the phantom. Notably, grey levels 2 and 5 are absent. 1 is the lowest grey level present in the ROI, and 6 the highest.

5.1.1 Computing image features

The digital phantom was designed to not require image processing prior to calculating the features. Thus, feature calculation is done directly on the phantom itself. The following should be taken into account for calculating image features:

• Discretisation is not required. All features are to be calculated using the phantom as it is. Al- ternatively, one could use a ﬁxed bin size discretisation of 1 or ﬁxed bin number discretisation of 6 bins, which does not alter the contents of the phantom.

• Grey level co-occurrence matrices are symmetrical and calculated for (Chebyshev) distance δ = 1.

144 CHAPTER 5. REFERENCE DATA SETS 145

• Neighbouring grey level dependence and neighbourhood grey tone diﬀerence matrices are likewise calculated for (Chebyshev) distance δ = 1. Additionally, the neighbouring grey level dependence coarseness parameter has the value α = 0.

• Because discretisation is lacking, most intensity-based statistical features will match their intensity histogram-based analogues in value.

• The ROI morphological and intensity masks are identical for the digital phantom, due to lack of re-segmentation.

5.2 Lung cancer CT image

A small data set of CT images from four non-small-cell lung carcinoma patients was made publicly available to serve as radiomics phantoms (DOI:10.17195/candat.2016.08.1). We use the image for the first patient (PAT1) to obtain feature reference values for different configurations of the image processing scheme, as detailed below. The CT image set is stored as a stack of slices in DICOM format. The image slices can be identified by the DCM IMG prefix. The gross tumour volume (GTV) was delineated and is used as the region of interest (ROI). Contour information is stored as an RT structure set in the DICOM file starting with DCM RS. For broader use, both the DICOM set and segmentation mask have been converted to the NIfTI format. When using the data in NIfTI format, both image stacks should be converted to (at least) 32-bit floating point and rounded to the nearest integer before further processing. We defined five image processing configurations to test different image processing methods, see Table 5.1. While most settings are self-explanatory, there are several aspects that require some attention. Configurations are divided in 2D and 3D approaches. For the 2D configurations (A, B), image interpolation is conducted within the slice, and likewise texture features are extracted from the in-slice plane, and not volumetrically (3D). For the 3D configurations (C-E) interpolation is conducted in three dimensions, and features are likewise extracted volumetrically. Discretisation is moreover required for texture, intensity histogram and intensity-volume histogram features, and both fixed bin number and fixed bin size algorithms are tested.

5.2.1 Notes on interpolation

Interpolation has a major influence on feature values. Different implementations of the same interpolation method may ostensibly provide the same functionality, but may use different interpolation grids. It is therefore recommended to read the documentation of the particular implementation to assess if the implementation allows or implements the following:

• The spatial origin of the original (input) grid in world coordinates matches the DICOM origin by deﬁnition.

• The size of the interpolation grid is determined by rounding the fractional grid size towards inﬁnity, i.e. a ceiling operation. This prevents the interpolation grid from disappearing for very small images, but is otherwise an arbitrary choice.

• The centers of the interpolation and original image grids should be aligned, i.e. the interpolation grid is centered on the center of the original image grid. This prevents spacing inconsistencies in the interpolation grid and avoids potential issues with grid orientation. CHAPTER 5. REFERENCE DATA SETS 146

x y 1 4 4 1 1 z 1 4 6 1 1

4 1 6 4 1 4 4 6 4 1

1 4 4 1 1

1 1 6 1 1

1 1 3 1 1

4 4 6 1 1

1 4 4 1 1

1 1 1 1 1

1 1 9 1 1 1 1 6 1 1

1 4 4 1 1

1 1 1 1 1

1 1 1 1 1 1 1 6 1 1

Figure 5.1 — Exploded view of the test volume. The number in each voxel corresponds with its grey level. Blue voxels are excluded from the region of interest. The coordinate system is so that x increases from left to right, y increases from back to front and z increases from top to bottom, as is indicated by the axis deﬁnition in the top-left. CHAPTER 5. REFERENCE DATA SETS 147

• The extent of the interpolation grid is, by deﬁnition, always equal or larger than that of the original grid. This means that intensities at the grid boundary are extrapolated. To facilitate this process, the image should be suﬃciently padded with voxels that take on the nearest boundary intensity.

• The floating point representation of the image and the ROI masks affects interpolation precision, and consequentially feature values. Image and ROI masks should at least be represented at full precision (32-bit) to avoid rounding errors. One example is the unintended exclusion of voxels from the interpolated ROI mask, which occurs when interpolation yields 0.4999. . . instead of 0.5. When images and ROI masks are converted to full precision from lower precision (e.g. 16-bit), values may require rounding if the original data were integer values, such as Hounsfield Units or the ROI mask labels.

More details are provided in Section 2.4.

5.2.2 Diagnostic features

Identifying issues with an implementation of the image processing sequence may be challenging. Multiple steps follow one another and diﬀerences propagate. Hence we deﬁne a small number of diagnostic features that describe how the image and ROI masks change with each image processing step. These diagnostic features also have reference values that may be found in IBSI compliance check spreadsheet.

Initial image stack. The following features may be used to describe the initial image stack (i.e. after loading image data for processing):

• Image dimensions. This describes the image dimensions in voxels along the diﬀerent image axes.

• Voxel dimensions. This describes the voxel dimensions in mm. The dimension along the z- axis is equal to the distance between the origin voxels of two adjacent slices, and is generally equal to the slice thickness.

• Mean intensity. This is the average intensity within the entire image.

• Minimum intensity. This is the lowest intensity within the entire image.

• Maximum intensity. This is the highest intensity within the entire image.

Interpolated image stack. The above features may also be used to describe the image stack after image interpolation.

Initial region of interest. The following descriptors are used to describe the region of interest (ROI) directly after segmentation of the image:

• ROI intensity mask dimensions. This describes the dimensions, in voxels, of the ROI intensity mask.

• ROI intensity mask bounding box dimensions. This describes the dimensions, in voxels, of the bounding box of the ROI intensity mask.

• ROI morphological mask bounding box dimensions. This describes the dimensions, in voxels, of the bounding box of the ROI morphological mask. CHAPTER 5. REFERENCE DATA SETS 148

• Number of voxels in the ROI intensity mask. This describes the number of voxels included in the ROI intensity mask.

• Number of voxels in the ROI morphological mask. This describes the number of voxels included in the ROI intensity mask.

• Mean ROI intensity. This is the mean intensity of image voxels within the ROI intensity mask.

• Minimum ROI intensity. This is the lowest intensity of image voxels within the ROI intensity mask.

• Maximum ROI intensity. This is the highest intensity of image voxels within the ROI intensity mask.

Interpolated region of interest. The same features can be used to describe the ROI after interpolation of the ROI mask.

Re-segmented region of interest. Again, the same features as above can be used to describe the ROI after re-segmentation.

5.2.3 Computing image features

Unlike the digital phantom, the lung cancer CT image does require additional image processing, which is done according to the processing conﬁgurations described in Table 5.1. The following should be taken into account when calculating image features:

• Grey level co-occurrence matrices are symmetrical and calculated for (Chebyshev) distance δ = 1.

• Intensity-based statistical features and their intensity histogram-based analogues will diﬀer in value due to discretisation, in contrast to the same features for the digital phantom.

• Due to re-segmentation, the ROI morphological and intensity masks are not identical.

• Calculation of IVH feature: since by default CT contains calibrated and discrete intensities, no separate discretisation prior to the calculation of intensity-volume histogram features is required. This is the case for configurations A, B and D (i.e. ‘calibrated intensity units – discrete case’). However, for configurations C and E, we re-discretise the ROI intensities prior to calculation of intensity-volume histogram features to allow for testing of of these methods. Configuration C simulates the ‘calibrated intensity units – continuous case’, while configuration E simulates the ‘arbitrary intensity units’ case where the re-segmentation range is not used. For details, please consult section 3.5. CHAPTER 5. REFERENCE DATA SETS 149 400] 2 , × 2 1000 5 × . σ − 3 2 2 × 2 5 0 × . σ 400] no [ 2 2 , × 2 1000 5 0 × . − 400] [ , 2 (axial) 2 500 5 0 × . − 400] [ , 500 − PAT1GTV-1 PAT1 GTV-1 PAT1 GTV-1 GTV-1 PAT1 GTV-1 PAT1 resampled voxel spacing (mm) 2 interpolation methodintensity roundingROI interpolation methodROI partial mask volume bilinear bilinear nearest integer 0 nearest trilinear integer trilinear nearest integer nearest integer trilinear trilinear tricubic spline trilinear range (HU)outlier filtering no [ no no 3 texture and IHIVHGLCM, NGTDM, NGLDM distanceGLSZM, GLDZM linkage 1 distanceNGLDM coarseness 1 FBS: 25 HU 1 FBN: 32 bins 1 FBS: 0.0 25 HU no FBN: 32 bins 1 0.0 FBN: 32 bins 1 no 1 0.0 FBS: 2.5 1 HU no 0.0 1 1 FBN: 1000 bins 0.0 Parametersample identifier Config. A Config. B Config. C Config. D Config. E ROI name slice-wise or single volume (3D)interpolation 2D 2D no 3D yes 3D yes 3D yes yes re-segmentation discretisation texture parameters Different configurations for image processing. For details, refer to the corresponding sections in chapter 2. ROI: region of interest; HU: Hounsfield Unit; Table 5.1 — IH: intensity histogram; FBS:grey fixed tone bin difference size; matrix; FBN: NGLDM: fixed neighbouring bin grey number; level dependence IVH: matrix; intensity-volume GLSZM: histogram; grey GLCM: level grey size level zone co-occurrence matrix; matrix; GLDZM: NGTDM: grey neighborhood level distance zone matrix. Appendix A

Digital phantom texture matrices

This section contains the texture matrices extracted from the digital phantom for reference purposes.

A.1 Grey level co-occurrence matrix (2D)

i j n i j n i j n i j n

1.0 1.0 10 1.0 1.0 16 1.0 1.0 18 1.0 1.0 20 1.0 4.0 4 1.0 4.0 2 1.0 4.0 2 1.0 4.0 2 4.0 1.0 4 3.0 6.0 2 4.0 1.0 2 1.0 6.0 1 4.0 4.0 6 4.0 1.0 2 4.0 1.0 2 (c) x: (0,1,0) 4.0 6.0 1 4.0 6.0 1 6.0 1.0 1 slice: 3 of 4 6.0 4.0 1 6.0 3.0 2 (d) x: (0,1,0) 6.0 6.0 4 6.0 4.0 1 slice: 4 of 4 (a) x: (0,1,0) (b) x: (0,1,0) slice: 1 of 4 slice: 2 of 4

i j n i j n i j n i j n

1.0 1.0 2 1.0 1.0 6 1.0 1.0 10 1.0 1.0 14 1.0 4.0 4 1.0 3.0 1 1.0 4.0 2 1.0 4.0 2 1.0 6.0 3 1.0 4.0 3 1.0 6.0 1 1.0 6.0 1 4.0 1.0 4 1.0 6.0 3 4.0 1.0 2 4.0 1.0 2 4.0 4.0 4 3.0 1.0 1 6.0 1.0 1 6.0 1.0 1 4.0 6.0 2 3.0 4.0 1 (g) x: (1,-1,0) (h) x: (1,-1,0) 6.0 1.0 3 4.0 1.0 3 slice: 3 of 4 slice: 4 of 4 6.0 4.0 2 4.0 3.0 1 6.0 1.0 3 (e) x: (1,-1,0) slice: 1 of 4 (f) x: (1,-1,0) slice: 2 of 4

150 APPENDIX A. DIGITAL PHANTOM TEXTURE MATRICES 151

i j n i j n i j n i j n

1.0 1.0 4 1.0 1.0 10 1.0 1.0 16 1.0 1.0 20 1.0 4.0 6 1.0 3.0 2 1.0 4.0 1 1.0 4.0 1 1.0 6.0 2 1.0 4.0 2 1.0 6.0 2 1.0 6.0 2 4.0 1.0 6 1.0 6.0 3 4.0 1.0 1 4.0 1.0 1 4.0 4.0 4 3.0 1.0 2 4.0 4.0 2 4.0 4.0 2 4.0 6.0 4 4.0 1.0 2 6.0 1.0 2 6.0 1.0 2 6.0 1.0 2 4.0 4.0 4 (k) d: (1,0,0) (l) d: (1,0,0) 6.0 4.0 4 4.0 6.0 1 slice: 3 of 4 slice: 4 of 4 6.0 1.0 3 (i) d: (1,0,0) 6.0 4.0 1 slice: 1 of 4 (j) d: (1,0,0) slice: 2 of 4

i j n i j n i j n i j n

1.0 1.0 6 1.0 1.0 10 1.0 1.0 12 1.0 1.0 16 1.0 4.0 3 1.0 3.0 2 1.0 4.0 2 1.0 4.0 2 1.0 6.0 1 1.0 4.0 1 1.0 6.0 1 1.0 6.0 1 4.0 1.0 3 1.0 6.0 2 4.0 1.0 2 4.0 1.0 2 4.0 4.0 2 3.0 1.0 2 6.0 1.0 1 6.0 1.0 1 4.0 6.0 4 4.0 1.0 1 (o) d: (1,1,0) (p) d: (1,1,0) 6.0 1.0 1 4.0 6.0 1 slice: 3 of 4 slice: 4 of 4 6.0 4.0 4 6.0 1.0 2 6.0 4.0 1 (m) d: (1,1,0) slice: 1 of 4 (n) d: (1,1,0) slice: 2 of 4 Table A.1 — Grey-level co-occurrence matrices extracted from the xy plane (2D) of the digital phantom using Chebyshev distance 1. x indicates the direction in (x, y, z) coordinates.

A.2 Grey level co-occurrence matrix (2D, merged) APPENDIX A. DIGITAL PHANTOM TEXTURE MATRICES 152

i j n i j n i j n i j n

1.0 1.0 22 1.0 1.0 42 1.0 1.0 56 1.0 1.0 70 1.0 4.0 17 1.0 3.0 5 1.0 4.0 7 1.0 4.0 7 1.0 6.0 6 1.0 4.0 8 1.0 6.0 4 1.0 6.0 5 4.0 1.0 17 1.0 6.0 8 4.0 1.0 7 4.0 1.0 7 4.0 4.0 16 3.0 1.0 5 4.0 4.0 2 4.0 4.0 2 4.0 6.0 11 3.0 4.0 1 6.0 1.0 4 6.0 1.0 5 6.0 1.0 6 3.0 6.0 2 (c) slice: 3 of 4 (d) slice: 4 of 4 6.0 4.0 11 4.0 1.0 8 6.0 6.0 4 4.0 3.0 1 4.0 4.0 4 (a) slice: 1 of 4 4.0 6.0 3 6.0 1.0 8 6.0 3.0 2 6.0 4.0 3

(b) slice: 2 of 4 Table A.2 — Merged grey-level co-occurrence matrices extracted from the xy plane (2D) of the digital phantom using Chebyshev distance 1.

A.3 Grey level co-occurrence matrix (3D)

i j n i j n i j n i j n

1.0 1.0 66 1.0 1.0 42 1.0 1.0 64 1.0 1.0 52 1.0 4.0 5 1.0 3.0 1 1.0 4.0 10 1.0 4.0 8 1.0 6.0 1 1.0 4.0 9 1.0 6.0 1 3.0 6.0 2 3.0 6.0 1 1.0 6.0 1 3.0 6.0 2 4.0 1.0 8 4.0 1.0 5 3.0 1.0 1 4.0 1.0 10 4.0 4.0 2 4.0 4.0 16 3.0 6.0 1 4.0 4.0 6 4.0 6.0 1 6.0 1.0 1 4.0 1.0 9 4.0 6.0 2 6.0 3.0 2 6.0 3.0 1 4.0 4.0 2 6.0 1.0 1 6.0 4.0 1 6.0 6.0 8 4.0 6.0 2 6.0 3.0 2 6.0 6.0 2 6.0 1.0 1 6.0 4.0 2 (a) x: (0,0,1) (d) x: (0,1,1) 6.0 3.0 1 6.0 6.0 4 6.0 4.0 2 (c) x: (0,1,0) 6.0 6.0 2

(b) x: (0,1,-1) APPENDIX A. DIGITAL PHANTOM TEXTURE MATRICES 153

i j n i j n i j n i j n

1.0 1.0 30 1.0 1.0 32 1.0 1.0 20 1.0 1.0 38 1.0 3.0 2 1.0 3.0 1 1.0 3.0 1 1.0 3.0 1 1.0 4.0 7 1.0 4.0 11 1.0 4.0 10 1.0 4.0 7 1.0 6.0 5 1.0 6.0 8 1.0 6.0 6 1.0 6.0 8 3.0 1.0 2 3.0 1.0 1 3.0 1.0 1 3.0 1.0 1 4.0 1.0 7 3.0 4.0 1 3.0 4.0 1 3.0 4.0 1 4.0 6.0 2 4.0 1.0 11 4.0 1.0 10 4.0 1.0 7 6.0 1.0 5 4.0 3.0 1 4.0 3.0 1 4.0 3.0 1 6.0 4.0 2 4.0 4.0 4 4.0 4.0 2 4.0 4.0 8 4.0 6.0 2 6.0 1.0 6 4.0 6.0 2 (e) x: (1,-1,-1) 6.0 1.0 8 6.0 1.0 8 (g) x: (1,-1,1) 6.0 4.0 2 6.0 4.0 2

(f) x: (1,-1,0) (h) x: (1,0,-1)

i j n i j n i j n i j n

1.0 1.0 50 1.0 1.0 34 1.0 1.0 32 1.0 1.0 44 1.0 3.0 2 1.0 3.0 2 1.0 3.0 1 1.0 3.0 2 1.0 4.0 10 1.0 4.0 8 1.0 4.0 6 1.0 4.0 8 1.0 6.0 9 1.0 6.0 7 1.0 6.0 4 1.0 6.0 5 3.0 1.0 2 3.0 1.0 2 3.0 1.0 1 3.0 1.0 2 4.0 1.0 10 4.0 1.0 8 3.0 4.0 1 4.0 1.0 8 4.0 4.0 12 4.0 4.0 8 4.0 1.0 6 4.0 4.0 2 4.0 6.0 5 4.0 6.0 3 4.0 3.0 1 4.0 6.0 5 6.0 1.0 9 6.0 1.0 7 4.0 6.0 3 6.0 1.0 5 6.0 4.0 5 6.0 4.0 3 6.0 1.0 4 6.0 4.0 5 6.0 4.0 3 (i) x: (1,0,0) (j) x: (1,0,1) (l) x: (1,1,0) (k) x: (1,1,-1)

i j n

1.0 1.0 32 1.0 3.0 1 1.0 4.0 6 1.0 6.0 6 3.0 1.0 1 3.0 4.0 1 4.0 1.0 6 4.0 3.0 1 4.0 4.0 2 4.0 6.0 1 6.0 1.0 6 6.0 4.0 1

(m) x: (1,1,1) Table A.3 — Grey-level co-occurrence matrices extracted volumetrically (3D) from the digital phantom using Chebyshev distance 1. x indicates the direction in (x, y, z) coordinates. APPENDIX A. DIGITAL PHANTOM TEXTURE MATRICES 154

A.4 Grey level co-occurrence matrix (3D, merged)

i j n

1.0 1.0 536 1.0 3.0 14 1.0 4.0 105 1.0 6.0 61 3.0 1.0 14 3.0 4.0 5 3.0 6.0 6 4.0 1.0 105 4.0 3.0 5 4.0 4.0 64 4.0 6.0 28 6.0 1.0 61 6.0 3.0 6 6.0 4.0 28 6.0 6.0 16

Table A.4 — Merged grey-level co-occurrence matrix extracted volumetrically (3D) from the digital phantom using Chebyshev distance 1.

A.5 Grey level run length matrix (2D)

i r n i r n i r n i r n

1.0 1.0 1.0 1.0 2.0 2.0 1.0 1.0 1.0 1.0 2.0 1.0 1.0 2.0 2.0 1.0 4.0 2.0 1.0 3.0 3.0 1.0 3.0 3.0 1.0 4.0 1.0 3.0 1.0 1.0 1.0 4.0 1.0 1.0 4.0 1.0 4.0 1.0 2.0 4.0 1.0 4.0 4.0 1.0 2.0 4.0 1.0 2.0 4.0 2.0 3.0 6.0 1.0 2.0 6.0 1.0 1.0 6.0 1.0 1.0 6.0 3.0 1.0 (b) x: (0,1,0) (c) x: (0,1,0) (d) x: (0,1,0) (a) x: (0,1,0) slice: 2 of 4 slice: 3 of 4 slice: 4 of 4 slice: 1 of 4 APPENDIX A. DIGITAL PHANTOM TEXTURE MATRICES 155

i r n i r n i r n i r n

1.0 1.0 7.0 1.0 1.0 6.0 1.0 1.0 5.0 1.0 1.0 3.0 1.0 2.0 1.0 1.0 2.0 3.0 1.0 2.0 3.0 1.0 2.0 3.0 4.0 1.0 5.0 3.0 1.0 1.0 1.0 3.0 1.0 1.0 3.0 2.0 4.0 3.0 1.0 4.0 1.0 4.0 4.0 1.0 2.0 4.0 1.0 2.0 6.0 1.0 3.0 6.0 1.0 2.0 6.0 1.0 1.0 6.0 1.0 1.0

(e) x: (1,-1,0) (f) x: (1,-1,0) (g) x: (1,-1,0) (h) x: (1,-1,0) slice: 1 of 4 slice: 2 of 4 slice: 3 of 4 slice: 4 of 4

i r n i r n i r n i r n

1.0 1.0 5.0 1.0 1.0 2.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 2.0 2.0 1.0 2.0 5.0 1.0 2.0 4.0 1.0 2.0 2.0 4.0 1.0 4.0 3.0 1.0 1.0 1.0 5.0 1.0 1.0 5.0 2.0 4.0 2.0 2.0 4.0 2.0 2.0 4.0 2.0 1.0 4.0 2.0 1.0 6.0 1.0 3.0 6.0 1.0 2.0 6.0 1.0 1.0 6.0 1.0 1.0

(i) x: (1,0,0) (j) x: (1,0,0) (k) x: (1,0,0) (l) x: (1,0,0) slice: 1 of 4 slice: 2 of 4 slice: 3 of 4 slice: 4 of 4

i r n i r n i r n i r n

1.0 1.0 3.0 1.0 1.0 2.0 1.0 1.0 3.0 1.0 1.0 2.0 1.0 2.0 3.0 1.0 2.0 5.0 1.0 2.0 4.0 1.0 2.0 3.0 4.0 1.0 6.0 3.0 1.0 1.0 1.0 3.0 1.0 1.0 3.0 1.0 4.0 2.0 1.0 4.0 1.0 4.0 4.0 1.0 2.0 1.0 4.0 1.0 6.0 1.0 3.0 6.0 1.0 2.0 6.0 1.0 1.0 4.0 1.0 2.0 6.0 1.0 1.0 (m) x: (1,1,0) (n) x: (1,1,0) (o) x: (1,1,0) slice: 1 of 4 slice: 2 of 4 slice: 3 of 4 (p) x: (1,1,0) slice: 4 of 4 Table A.5 — Grey-level run length matrices extracted from the xy plane (2D) of the digital phantom. x indicates the direction in (x, y, z) coordinates. APPENDIX A. DIGITAL PHANTOM TEXTURE MATRICES 156

A.6 Grey level run length matrix (2D, merged)

i r n i r n i r n i r n

1.0 1.0 16.0 1.0 1.0 10.0 1.0 1.0 10.0 1.0 1.0 6.0 1.0 2.0 8.0 1.0 2.0 15.0 1.0 2.0 11.0 1.0 2.0 9.0 1.0 4.0 1.0 1.0 4.0 2.0 1.0 3.0 5.0 1.0 3.0 6.0 4.0 1.0 17.0 3.0 1.0 4.0 1.0 4.0 1.0 1.0 4.0 2.0 4.0 2.0 6.0 4.0 1.0 12.0 1.0 5.0 1.0 1.0 5.0 2.0 4.0 3.0 1.0 4.0 2.0 2.0 4.0 1.0 6.0 4.0 1.0 6.0 6.0 1.0 9.0 6.0 1.0 8.0 4.0 2.0 1.0 4.0 2.0 1.0 6.0 3.0 1.0 6.0 1.0 4.0 6.0 1.0 4.0 (b) slice: 2 of 4 (a) slice: 1 of 4 (c) slice: 3 of 4 (d) slice: 4 of 4 Table A.6 — Merged grey-level run length matrices extracted from the xy plane (2D) of the digital phantom.

A.7 Grey level run length matrix (3D)

i r n i r n i r n i r n

1.0 1.0 1.0 1.0 1.0 11.0 1.0 1.0 2.0 1.0 1.0 10.0 1.0 2.0 6.0 1.0 2.0 15.0 1.0 2.0 5.0 1.0 2.0 5.0 1.0 3.0 3.0 1.0 3.0 3.0 1.0 3.0 6.0 1.0 3.0 6.0 1.0 4.0 7.0 3.0 1.0 1.0 1.0 4.0 5.0 1.0 4.0 3.0 3.0 1.0 1.0 4.0 1.0 14.0 3.0 1.0 1.0 3.0 1.0 1.0 4.0 1.0 4.0 4.0 2.0 1.0 4.0 1.0 10.0 4.0 1.0 14.0 4.0 2.0 2.0 6.0 1.0 5.0 4.0 2.0 3.0 4.0 2.0 1.0 4.0 4.0 2.0 6.0 2.0 1.0 6.0 1.0 4.0 6.0 1.0 5.0 6.0 1.0 1.0 6.0 3.0 1.0 6.0 2.0 1.0 (b) x: (0,1,-1) 6.0 2.0 1.0 (c) x: (0,1,0) (d) x: (0,1,1) 6.0 4.0 1.0

(a) x: (0,0,1) APPENDIX A. DIGITAL PHANTOM TEXTURE MATRICES 157

i r n i r n i r n i r n

1.0 1.0 22.0 1.0 1.0 21.0 1.0 1.0 30.0 1.0 1.0 16.0 1.0 2.0 11.0 1.0 2.0 10.0 1.0 2.0 10.0 1.0 2.0 12.0 1.0 3.0 2.0 1.0 3.0 3.0 3.0 1.0 1.0 1.0 3.0 2.0 3.0 1.0 1.0 3.0 1.0 1.0 4.0 1.0 14.0 1.0 4.0 1.0 4.0 1.0 16.0 4.0 1.0 13.0 4.0 2.0 1.0 3.0 1.0 1.0 6.0 1.0 7.0 4.0 3.0 1.0 6.0 1.0 7.0 4.0 1.0 8.0 6.0 1.0 7.0 4.0 2.0 4.0 (e) x: (1,-1,-1) (g) x: (1,-1,1) 6.0 1.0 7.0 (f) x: (1,-1,0) (h) x: (1,0,-1)

i r n i r n i r n i r n

1.0 1.0 9.0 1.0 1.0 19.0 1.0 1.0 20.0 1.0 1.0 10.0 1.0 2.0 13.0 1.0 2.0 12.0 1.0 2.0 12.0 1.0 2.0 15.0 1.0 5.0 3.0 1.0 3.0 1.0 1.0 3.0 2.0 1.0 3.0 2.0 3.0 1.0 1.0 1.0 4.0 1.0 3.0 1.0 1.0 1.0 4.0 1.0 4.0 1.0 4.0 3.0 1.0 1.0 4.0 1.0 16.0 3.0 1.0 1.0 4.0 2.0 6.0 4.0 1.0 8.0 6.0 1.0 7.0 4.0 1.0 14.0 6.0 1.0 7.0 4.0 2.0 4.0 4.0 2.0 1.0 (k) x: (1,1,-1) 6.0 1.0 7.0 6.0 1.0 7.0 (i) x: (1,0,0) (j) x: (1,0,1) (l) x: (1,1,0)

i r n

1.0 1.0 19.0 1.0 2.0 14.0 1.0 3.0 1.0 3.0 1.0 1.0 4.0 1.0 14.0 4.0 2.0 1.0 6.0 1.0 7.0

(m) x: (1,1,1) Table A.7 — Grey-level run length matrices extracted volumetrically (3D) from the digital phantom. x indicates the direction in (x, y, z) coordinates. APPENDIX A. DIGITAL PHANTOM TEXTURE MATRICES 158

A.8 Grey level run length matrix (3D, merged)

i r n

1.0 1.0 190.0 1.0 2.0 140.0 1.0 3.0 31.0 1.0 4.0 18.0 1.0 5.0 3.0 3.0 1.0 13.0 4.0 1.0 149.0 4.0 2.0 24.0 4.0 3.0 1.0 4.0 4.0 2.0 6.0 1.0 78.0 6.0 2.0 3.0 6.0 3.0 1.0 6.0 4.0 1.0

Table A.8 — Merged grey-level run length matrix extracted volumetrically (3D) from the digital phantom.

A.9 Grey level size zone matrix (2D)

i s n i s n i s n i s n

1.0 3 1 1.0 4 1 1.0 14 1 1.0 15 1 1.0 6 1 1.0 8 1 4.0 2 1 4.0 2 1 4.0 2 1 3.0 1 1 6.0 1 1 6.0 1 1 4.0 6 1 4.0 2 2 (c) slice: 3 of 4 (d) slice: 4 of 4 6.0 3 1 6.0 1 2

(a) slice: 1 of 4 (b) slice: 2 of 4 Table A.9 — Grey level size zone matrices extracted from the xy plane (2D) of the digital phantom.

A.10 Grey level size zone matrix (3D)

i s n

1.0 50 1 3.0 1 1 4.0 2 1 4.0 14 1 6.0 7 1

Table A.10 — Grey level size zone matrix extracted volumetrically (3D) from the digital phantom. APPENDIX A. DIGITAL PHANTOM TEXTURE MATRICES 159

A.11 Grey level distance zone matrix (2D)

i d n i d n i d n i d n

1.0 1.0 2 1.0 1.0 2 1.0 1.0 1 1.0 1.0 1 4.0 1.0 2 3.0 2.0 1 4.0 1.0 1 4.0 1.0 1 6.0 1.0 1 4.0 1.0 2 6.0 1.0 1 6.0 1.0 1 6.0 1.0 1 (a) slice: 1 of 4 (c) slice: 3 of 4 (d) slice: 4 of 4 6.0 2.0 1

(b) slice: 2 of 4 Table A.11 — Grey level distance zone matrices extracted from the xy plane (2D) of the digital phantom.

A.12 Grey level distance zone matrix (3D)

i d n

1.0 1.0 1 3.0 1.0 1 4.0 1.0 2 6.0 1.0 1

Table A.12 — Grey level distance zone matrix extracted volumetrically (3D) from the digital phantom.

A.13 Neighbourhood grey tone diﬀerence matrix (2D)

i s n i s n i s n i s n

1.0 14.575 9 1.0 11.928571 12 1.0 7.985714 14 1.0 7.582143 15 4.0 5.775 8 3.0 0.375000 1 4.0 4.650000 2 4.0 4.650000 2 6.0 7.325 3 4.0 4.800000 4 6.0 5.000000 1 6.0 5.000000 1 6.0 8.000000 2 (a) slice: 1 of 4 (c) slice: 3 of 4 (d) slice: 4 of 4 (b) slice: 2 of 4 Table A.13 — Neighbourhood grey tone diﬀerence matrices extracted from the xy plane (2D) of the digital phantom using Chebyshev distance 1.

A.14 Neighbourhood grey tone diﬀerence matrix (3D) APPENDIX A. DIGITAL PHANTOM TEXTURE MATRICES 160

i s n

1.0 39.946954 50 3.0 0.200000 1 4.0 20.825401 16 6.0 24.127005 7

Table A.14 — Neighbourhood grey tone diﬀerence matrix extracted volumetrically (3D) from the digital phantom using Chebyshev distance 1.

A.15 Neighbouring grey level dependence matrix (2D)

i j s i j s i j s i j s

1.0 2.0 3 1.0 3.0 2 1.0 3.0 1 1.0 3.0 1 1.0 3.0 1 1.0 4.0 6 1.0 4.0 5 1.0 4.0 3 1.0 4.0 3 1.0 6.0 4 1.0 5.0 3 1.0 5.0 3 1.0 5.0 2 3.0 1.0 1 1.0 6.0 3 1.0 6.0 4 4.0 2.0 2 4.0 2.0 4 1.0 7.0 2 1.0 7.0 1 4.0 3.0 4 6.0 1.0 2 4.0 2.0 2 1.0 8.0 3 4.0 4.0 2 6.0 1.0 1 4.0 2.0 2 (b) slice: 2 of 4 6.0 2.0 2 6.0 1.0 1 (c) slice: 3 of 4 6.0 3.0 1 (d) slice: 4 of 4 (a) slice: 1 of 4 Table A.15 — Neighbouring grey level dependence matrices extracted from the xy plane (2D) of the digital phantom using Chebyshev distance 1 and coarseness 0.

A.16 Neighbouring grey level dependence matrix (3D) APPENDIX A. DIGITAL PHANTOM TEXTURE MATRICES 161

i j s

1.0 5.0 2 1.0 6.0 2 1.0 7.0 1 1.0 8.0 6 1.0 9.0 4 1.0 10.0 6 1.0 11.0 5 1.0 12.0 5 1.0 13.0 3 1.0 14.0 2 1.0 15.0 5 1.0 16.0 3 1.0 17.0 3 1.0 18.0 2 1.0 21.0 1 3.0 1.0 1 4.0 2.0 2 4.0 4.0 2 4.0 5.0 6 4.0 6.0 4 4.0 7.0 2 6.0 2.0 1 6.0 3.0 4 6.0 4.0 1 6.0 5.0 1

Table A.16 — Neighbouring grey level dependence matrix extracted volumetrically (3D) from the digital phantom using Chebyshev distance 1 and coarseness 0. Bibliography

[1] H. J. W. L. Aerts, E. Rios-Velazquez, R. T. H. Leijenaar, C. Parmar, P. Grossmann, S. Cavalho, J. Bussink, R. Monshouwer, B. Haibe-Kains, D. Rietveld, F. J. P. Hoebers, M. M. Rietbergen, C. R. Leemans, A. Dekker, J. Quackenbush, R. J. Gillies, and P. Lambin. De- coding tumour phenotype by noninvasive imaging using a quantitative radiomics approach. Nature communications, 5:4006, 2014.

[2] S. D. Ahipa¸sao˘glu. Fast algorithms for the minimum volume estimator. Journal of Global Optimization, 62(2):351–370, 2015.

[3] F. Albregtsen, B. Nielsen, and H. Danielsen. Adaptive gray level run length features from class distance matrices. In Proceedings 15th International Conference on Pattern Recognition. ICPR-2000, volume 3, pages 738–741. IEEE Comput. Soc, 2000.

[4] B. A. Altazi, G. G. Zhang, D. C. Fernandez, M. E. Montejo, D. Hunt, J. Werner, M. C. Biagioli, and E. G. Moros. Reproducibility of F18-FDG PET radiomic features for diﬀerent cervical tumor segmentation methods, gray-level discretization, and reconstruction algorithms. Journal of applied clinical medical physics, 18(6):32–48, 2017.

[5] M. Amadasun and R. King. Textural features corresponding to textural properties. IEEE Transactions on Systems, Man and Cybernetics, 19(5):1264–1273, 1989.

[6] I. Apostolova, I. G. Steﬀen, F. Wedel, A. Lougovski, S. Marnitz, T. Derlin, H. Amthauer, R. Buchert, F. Hofheinz, and W. Brenner. Asphericity of pretherapeutic tumour FDG uptake provides independent prognostic value in head-and-neck cancer. European radiology, 24(9): 2077–87, 2014.

[7] J. Atkinson A.J., W. Colburn, V. DeGruttola, D. DeMets, G. Downing, D. Hoth, J. Oates, C. Peck, R. Schooley, B. Spilker, J. Woodcock, and S. Zeger. Biomarkers and surrogate en- dpoints: Preferred deﬁnitions and conceptual framework. Clinical Pharmacology and Thera- peutics, 69(3):89–95, 2001.

[8] C. Bailly, C. Bodet-Milin, S. Couespel, H. Necib, F. Kraeber-Bod´er´e, C. Ansquer, and T. Carlier. Revisiting the Robustness of PET-Based Textural Features in the Context of Multi-Centric Trials. PloS one, 11(7):e0159984, 2016.

[9] M. A. Balafar, A. R. Ramli, M. I. Saripan, and S. Mashohor. Review of brain MRI image segmentation methods. Artiﬁcial Intelligence Review, 33(3):261–274, 2010.

[10] G. Barequet and S. Har-Peled. Eﬃciently Approximating the Minimum-Volume Bounding Box of a Point Set in Three Dimensions. Journal of Algorithms, 38(1):91–109, 2001.

[11] R. Boellaard, R. Delgado-Bolton, W. J. G. Oyen, F. Giammarile, K. Tatsch, W. Eschner, F. J. Verzijlbergen, S. F. Barrington, L. C. Pike, W. A. Weber, S. G. Stroobants, D. Delbeke, K. J. Donohoe, S. Holbrook, M. M. Graham, G. Testanera, O. S. Hoekstra, J. M. Zijlstra, E. P. Visser, C. J. Hoekstra, J. Pruim, A. T. Willemsen, B. Arends, J. Kotzerke, A. Bockisch,

162 BIBLIOGRAPHY 163

T. Beyer, A. Chiti, and B. J. Krause. FDG PET/CT: EANM procedure guidelines for tumour imaging: version 2.0. European journal of nuclear medicine and molecular imaging, 42(2):328– 54, 2015.

[12] N. Boussion, C. C. Le Rest, M. Hatt, and D. Visvikis. Incorporation of wavelet-based denoising in iterative deconvolution for partial volume correction in whole-body PET imaging. European journal of nuclear medicine and molecular imaging, 36(7):1064–75, 2009.

[13] J. C. Caicedo, S. Cooper, F. Heigwer, S. Warchal, P. Qiu, C. Molnar, A. S. Vasilevich, J. D. Barry, H. S. Bansal, O. Kraus, M. Wawer, L. Paavolainen, M. D. Herrmann, M. Rohban, J. Hung, H. Hennig, J. Concannon, I. Smith, P. A. Clemons, S. Singh, P. Rees, P. Horvath, R. G. Linington, and A. E. Carpenter. Data-analysis strategies for image-based cell proﬁling. Nature Methods, 14(9):849–863, 2017.

[14] C. Chan and S. Tan. Determination of the minimum bounding box of an arbitrary solid: an iterative approach. Computers and Structures, 79(15):1433–1449, 2001.

[15] A. Chu, C. M. Sehgal, and J. F. Greenleaf. Use of gray value distribution of run lengths for texture analysis. Pattern Recognition Letters, 11(6):415–419, 1990.

[16] K. Clark, B. Vendt, K. Smith, J. Freymann, J. Kirby, P. Koppel, S. Moore, S. Phillips, D. Maﬃtt, M. Pringle, L. Tarbox, and F. Prior. The Cancer Imaging Archive (TCIA): maintaining and operating a public information repository. Journal of digital imaging, 26 (6):1045–57, 2013.

[17] L. P. Clarke, R. J. Nordstrom, H. Zhang, P. Tandon, Y. Zhang, G. Redmond, K. Farahani, G. Kelloﬀ, L. Henderson, L. Shankar, J. Deye, J. Capala, and P. Jacobs. The Quantitative Imaging Network: NCI’s Historical Perspective and Planned Goals. Translational oncology, 7 (1):1–4, 2014.

[18] D. A. Clausi. An analysis of co-occurrence texture statistics as a function of grey level quantization. Canadian Journal of Remote Sensing, 28(1):45–62, 2002.

[19] G. Collewet, M. Strzelecki, and F. Mariette. Inﬂuence of MRI acquisition protocols and image intensity normalization methods on texture classiﬁcation. Magnetic resonance imaging, 22(1): 81–91, 2004.

[20] E. C. Da Silva, A. C. Silva, A. C. De Paiva, and R. A. Nunes. Diagnosis of lung nodule using Moran’s index and Geary’s coeﬃcient in computerized tomography images. Pattern Analysis and Applications, 11(1):89–99, 2008.

[21] M. R. T. Dale, P. Dixon, M.-J. Fortin, P. Legendre, D. E. Myers, and M. S. Rosenberg. Conceptual and mathematical relationships among methods for spatial analysis. Ecography, 25(5):558–577, 2002.

[22] B. V. Dasarathy and E. B. Holder. Image characterizations based on joint gray level-run length distributions. Pattern Recognition Letters, 12(8):497–502, 1991.

[23] A. Depeursinge and J. Fageot. Biomedical Texture Operators and Aggregation Functions. In A. Depeursinge, J. Fageot, and O. Al-Kadi, editors, Biomedical texture analysis, chapter 3, pages 63–101. Academic Press, London, UK, 1st edition, 2017.

[24] A. Depeursinge, A. Foncubierta-Rodriguez, D. Van De Ville, and H. M¨uller.Three-dimensional solid texture analysis in biomedical imaging: review and opportunities. Medical image analysis, 18(1):176–96, 2014. BIBLIOGRAPHY 164

[25] M.-C. Desseroit, F. Tixier, W. A. Weber, B. A. Siegel, C. Cheze Le Rest, D. Visvikis, and M. Hatt. Reliability of PET/CT Shape and Heterogeneity Features in Functional and Mor- phologic Components of Non-Small Cell Lung Cancer Tumors: A Repeatability Analysis in a Prospective Multicenter Cohort. Journal of nuclear medicine, 58(3):406–411, 2017.

[26] I. El Naqa. Image Processing and Analysis of PET and Hybrid PET Imaging, pages 285–301. Springer International Publishing, Cham, 2017.

[27] I. El Naqa, P. W. Grigsby, A. Apte, E. Kidd, E. Donnelly, D. Khullar, S. Chaudhari, D. Yang, M. Schmitt, R. Laforest, W. L. Thorstad, and J. O. Deasy. Exploring feature-based approaches in PET images for predicting cancer treatment outcomes. Pattern recognition, 42(6):1162– 1171, 2009.

[28] European Society of Radiology (ESR). ESR statement on the stepwise development of imaging biomarkers. Insights into imaging, 4(2):147–52, 2013.

[29] V. Frings, F. H. P. van Velden, L. M. Velasquez, W. Hayes, P. M. van de Ven, O. S. Hoekstra, and R. Boellaard. Repeatability of metabolically active tumor volume measurements with FDG PET/CT in advanced gastrointestinal malignancies: a multicenter study. Radiology, 273(2):539–48, 2014.

[30] M. M. Galloway. Texture analysis using gray level run lengths. Computer Graphics and Image Processing, 4(2):172–179, 1975.

[31] R. C. Geary. The Contiguity Ratio and Statistical Mapping. The Incorporated Statistician, 5 (3):115–145, 1954.

[32] L. Gjesteby, B. De Man, Y. Jin, H. Paganetti, J. Verburg, D. Giantsoudi, and G. Wang. Metal Artifact Reduction in CT: Where Are We After Four Decades? IEEE Access, 4:5826–5849, 2016.

[33] H. Gudbjartsson and S. Patz. The Rician distribution of noisy MRI data. Magnetic resonance in medicine, 34(6):910–4, 1995.

[34] E. L. Hall, R. P. Kruger, J. Samuel, D. Dwyer, R. W. McLaren, D. L. Hall, and G. Lodwick. A Survey of Preprocessing and Feature Extraction Techniques for Radiographic Images. IEEE Transactions on Computers, C-20(9):1032–1044, 1971.

[35] R. M. Haralick. Statistical and structural approaches to texture. Proceedings of the IEEE, 67 (5):786–804, 1979.

[36] R. M. Haralick, K. Shanmugam, and I. Dinstein. Textural Features for Image Classiﬁcation. IEEE Transactions on Systems, Man, and Cybernetics, 3(6):610–621, 1973.

[37] M. Hatt, M. Majdoub, M. Valli`eres, F. Tixier, C. C. Le Rest, D. Groheux, E. Hindi´e, A. Mar- tineau, O. Pradier, R. Hustinx, R. Perdrisot, R. Guillevin, I. El Naqa, and D. Visvikis. 18F- FDG PET uptake characterization through texture analysis: investigating the complementary nature of heterogeneity and functional tumor volume in a multi-cancer site patient cohort. Journal of nuclear medicine, 56(1):38–44, 2015.

[38] M. Hatt, F. Tixier, L. Pierce, P. E. Kinahan, C. C. Le Rest, and D. Visvikis. Characterization of PET/CT images using texture analysis: the past, the present. . . any future? European journal of nuclear medicine and molecular imaging, 44(1):151–165, 2017.

[39] R. M. Heiberger and B. Holland. Statistical Analysis and Data Display. Springer Texts in Statistics. Springer New York, New York, NY, 2015. BIBLIOGRAPHY 165

[40] L. G. Khachiyan. Rounding of Polytopes in the Real Number Model of Computation. Math- ematics of Operations Research, 21(2):307–320, 1996.

[41] P. Lambin, R. T. H. Leijenaar, T. M. Deist, J. Peerlings, E. E. C. de Jong, J. van Timmeren, S. Sanduleanu, R. T. H. M. Larue, A. J. G. Even, A. Jochems, Y. van Wijk, H. Woodruﬀ, J. van Soest, T. Lustberg, E. Roelofs, W. J. C. van Elmpt, A. L. A. J. Dekker, F. M. Mottaghy, J. E. Wildberger, and S. Walsh. Radiomics: the bridge between medical imaging and personalized medicine. Nature reviews. Clinical oncology, 14(12):749–762, 2017.

[42] R. T. H. M. Larue, J. E. van Timmeren, E. E. C. de Jong, G. Feliciani, R. T. H. Leijenaar, W. M. J. Schreurs, M. N. Sosef, F. H. P. J. Raat, F. H. R. van der Zande, M. Das, W. J. C. van Elmpt, and P. Lambin. Inﬂuence of gray level discretization on radiomic feature stability for diﬀerent CT scanners, tube currents and slice thicknesses: a comprehensive phantom study. Acta oncologica, pages 1–10, 2017.

[43] A. Le Pogam, H. Hanzouli, M. Hatt, C. Cheze Le Rest, and D. Visvikis. Denoising of PET images by combining wavelets and curvelets for improved preservation of resolution and quant- itation. Medical image analysis, 17(8):877–91, 2013.

[44] R. T. H. Leijenaar, G. Nalbantov, S. Carvalho, W. J. C. van Elmpt, E. G. C. Troost, R. Boel- laard, H. J. W. L. Aerts, R. J. Gillies, and P. Lambin. The eﬀect of SUV discretization in quantitative FDG-PET Radiomics: the need for standardized methodology in tumor texture analysis. Scientiﬁc reports, 5(August):11075, 2015.

[45] T. Lewiner, H. Lopes, A. W. Vieira, and G. Tavares. Eﬃcient Implementation of Marching Cubes’ Cases with Topological Guarantees. Journal of Graphics Tools, 8(2):1–15, 2003.

[46] E. J. Limkin, S. Reuzé,A. Carré,R. Sun, A. Schernberg, A. Alexis, E. Deutsch, C. Ferté,and C. Robert. The complexity of tumor shape, spiculatedness, correlates with tumor radiomic shape features. Sci. Rep., 9(1):4329, 2019.

[47] S. P. Lloyd. Least Squares Quantization in PCM. IEEE Transactions on Information Theory, 28(2):129–137, 1982.

[48] W. E. Lorensen and H. E. Cline. Marching cubes: A high resolution 3D surface construction algorithm. ACM SIGGRAPH Computer Graphics, 21(4):163–169, 1987.

[49] D. Mackin, X. Fave, L. Zhang, J. Yang, A. K. Jones, C. S. Ng, and L. Court. Harmonizing the pixel size in retrospective computed tomography radiomics studies. PLOS ONE, 12(9): e0178524, 2017.

[50] J. Max. Quantizing for minimum distortion. IEEE Transactions on Information Theory, 6 (1):7–12, 1960.

[51] M. A. Mazurowski, N. M. Czarnek, L. M. Collins, K. B. Peters, and K. Clark. Predicting outcomes in glioblastoma patients using computerized analysis of tumor shape: preliminary data. In G. D. Tourassi and S. G. Armato, editors, SPIE Medical Imaging, volume 9785, page 97852T, 2016.

[52] P. A. P. Moran. Notes on continuous stochastic phenomena. Biometrika, 37:17–23, 1950.

[53] J. L. Mulshine, D. S. Gierada, S. G. Armato, R. S. Avila, D. F. Yankelevitz, E. A. Kazerooni, M. F. McNitt-Gray, A. J. Buckler, and D. C. Sullivan. Role of the Quantitative Imaging Biomarker Alliance in optimizing CT for the evaluation of lung cancer screen-detected nodules. Journal of the American College of Radiology, 12(4):390–5, 2015.

[54] R. J. Nordstrom. The quantitative imaging network in precision medicine. Tomography, 2(4): 239, 2016. BIBLIOGRAPHY 166

[55] J. P. B. O’Connor, E. O. Aboagye, J. E. Adams, H. J. W. L. Aerts, S. F. Barrington, A. J. Beer, R. Boellaard, S. E. Bohndiek, M. Brady, G. Brown, D. L. Buckley, T. L. Chenevert, L. P. Clarke, S. Collette, G. J. Cook, N. M. DeSouza, J. C. Dickson, C. Dive, J. L. Evelhoch, C. Faivre-Finn, F. A. Gallagher, F. J. Gilbert, R. J. Gillies, V. Goh, J. R. Griﬃths, A. M. Groves, S. Halligan, A. L. Harris, D. J. Hawkes, O. S. Hoekstra, E. P. Huang, B. F. Hutton, E. F. Jackson, G. C. Jayson, A. Jones, D.-M. Koh, D. Lacombe, P. Lambin, N. Lassau, M. O. Leach, T.-Y. Lee, E. L. Leen, J. S. Lewis, Y. Liu, M. F. Lythgoe, P. Manoharan, R. J. Maxwell, K. A. Miles, B. Morgan, S. Morris, T. Ng, A. R. Padhani, G. J. M. Parker, M. Partridge, A. P. Pathak, A. C. Peet, S. Punwani, A. R. Reynolds, S. P. Robinson, L. K. Shankar, R. A. Sharma, D. Soloviev, S. Stroobants, D. C. Sullivan, S. A. Taylor, P. S. Tofts, G. M. Tozer, M. van Herk, S. Walker-Samuel, J. Wason, K. J. Williams, P. Workman, T. E. Yankeelov, K. M. Brindle, L. M. McShane, A. Jackson, and J. C. Waterton. Imaging biomarker roadmap for cancer studies. Nature Reviews Clinical Oncology, 14(3):169–186, 2017.

[56] J. O’Rourke. Finding minimal enclosing boxes. International Journal of Computer and In- formation Sciences, 14(3):183–199, 1985.

[57] S. Sanduleanu, H. C. Woodruﬀ, E. E. C. de Jong, J. E. van Timmeren, A. Jochems, L. Dubois, and P. Lambin. Tracking tumor biology with radiomics: A systematic review utilizing a radiomics quality score. Radiother. Oncol., 127(3):349–360, 2018.

[58] S. Schirra. How Reliable Are Practical Point-in-Polygon Strategies? In Algorithms - ESA 2008, pages 744–755. Springer Berlin Heidelberg, Berlin, Heidelberg, 2008.

[59] M. Shaﬁq-Ul-Hassan, G. G. Zhang, K. Latiﬁ, G. Ullah, D. C. Hunt, Y. Balagurunathan, M. A. Abdalah, M. B. Schabath, D. G. Goldgof, D. Mackin, L. E. Court, R. J. Gillies, and E. G. Moros. Intrinsic dependencies of CT radiomic features on voxel size and number of gray levels. Medical physics, 44(3):1050–1062, 2017.

[60] I. Shiri, A. Rahmim, P. Ghaﬀarian, P. Geramifar, H. Abdollahi, and A. Bitarafan-Rajabi. The impact of image reconstruction settings on 18F-FDG PET radiomic features: multi-scanner phantom and patient studies. European Radiology, 27(11):4498–4509, 2017.

[61] J. G. Sled, A. P. Zijdenbos, and A. C. Evans. A nonparametric method for automatic correction of intensity nonuniformity in MRI data. IEEE transactions on medical imaging, 17(1):87–97, 1998.

[62] K. Smith, Y. Li, F. Piccinini, G. Csucs, C. Balazs, A. Bevilacqua, and P. Horvath. CIDRE: An illumination-correction method for optical microscopy. Nature Methods, 12(5):404–406, 2015.

[63] L.-K. Soh and C. Tsatsoulis. Texture analysis of sar sea ice imagery using gray level co- occurrence matrices. IEEE Transactions on Geoscience and Remote Sensing, 37(2):780–795, 1999.

[64] M. Sollini, L. Cozzi, L. Antunovic, A. Chiti, and M. Kirienko. PET Radiomics in NSCLC: state of the art and a proposal for harmonization of methodology. Scientiﬁc reports, 7(1):358, 2017.

[65] C. Solomon and T. Breckon. Features. In Fundamentals of Digital Image Processing, chapter 9, pages 235–262. John Wiley & Sons, Ltd, Chichester, UK, 2011.

[66] M. Soret, S. L. Bacharach, and I. Buvat. Partial-volume eﬀect in PET tumor imaging. Journal of nuclear medicine, 48(6):932–45, 2007. BIBLIOGRAPHY 167

[67] P. Stelldinger, L. J. Latecki, and M. Siqueira. Topological equivalence between a 3D object and the reconstruction of its digital image. IEEE transactions on pattern analysis and machine intelligence, 29(1):126–40, 2007.

[68] D. C. Sullivan, N. A. Obuchowski, L. G. Kessler, D. L. Raunig, C. Gatsonis, E. P. Huang, M. Kondratovich, L. M. McShane, A. P. Reeves, D. P. Barboriak, A. R. Guimaraes, R. L. Wahl, and RSNA-QIBA Metrology Working Group. Metrology Standards for Quantitative Imaging Biomarkers. Radiology, 277(3):813–25, 2015.

[69] C. Sun and W. G. Wee. Neighboring gray level dependence matrix for texture classiﬁcation. Computer Vision, Graphics, and Image Processing, 23(3):341–352, 1983.

[70] P. Th´evenaz, T. Blu, and M. Unser. Image interpolation and resampling. In Handbook of medical imaging, pages 393–420. Academic Press, Inc., 2000.

[71] G. Thibault, J. Angulo, and F. Meyer. Advanced statistical matrices for texture characterization: application to cell classiﬁcation. IEEE transactions on bio-medical engineering, 61(3): 630–7, 2014.

[72] M. J. Todd and E. A. Yldrm. On Khachiyan’s algorithm for the computation of minimum- volume enclosing ellipsoids. Discrete Applied Mathematics, 155(13):1731–1744, 2007.

[73] A. Traverso, L. Wee, A. Dekker, and R. Gillies. Repeatability and reproducibility of radiomic features: A systematic review. Int. J. Radiat. Oncol. Biol. Phys., 102(4):1143–1158, 2018.

[74] M. Unser. Sum and diﬀerence histograms for texture classiﬁcation. IEEE transactions on pattern analysis and machine intelligence, 8(1):118–125, 1986.

[75] M. Vaidya, K. M. Creach, J. Frye, F. Dehdashti, J. D. Bradley, and I. El Naqa. Combined PET/CT image characteristics for radiotherapy tumor response in lung cancer. Radiotherapy and oncology, 102(2):239–45, 2012.

[76] M. Valli`eres,C. R. Freeman, S. R. Skamene, and I. El Naqa. A radiomics model from joint FDG-PET and MRI texture features for the prediction of lung metastases in soft-tissue sarcomas of the extremities. Physics in medicine and biology, 60(14):5471–96, 2015.

[77] M. Valli`eres,C. R. Freeman, S. R. Skamene, and I. El Naqa. Data from: A radiomics model from joint FDG-PET and MRI texture features for the prediction of lung metastases in soft- tissue sarcomas of the extremities, 2015.

[78] M. Valli`eres,E. Kay-Rivest, L. J. Perrin, X. Liem, C. Furstoss, H. J. W. L. Aerts, N. Kha- ouam, P. F. Nguyen-Tan, C.-S. Wang, K. Sultanem, J. Seuntjens, and I. El Naqa. Radiomics strategies for risk assessment of tumour failure in head-and-neck cancer. Scientiﬁc reports, 7: 10117, 2017.

[79] M. Vallieres, A. Zwanenburg, B. Badic, C. Cheze-Le Rest, D. Visvikis, and M. Hatt. Respons- ible radiomics research for faster clinical translation, 2017.

[80] L. V. van Dijk, C. L. Brouwer, A. van der Schaaf, J. G. Burgerhof, R. J. Beukinga, J. A. Langendijk, N. M. Sijtsema, and R. J. Steenbakkers. CT image biomarkers to improve patient-speciﬁc prediction of radiation-induced xerostomia and sticky saliva. Radiotherapy and Oncology, 122(2):185–191, 2017.

[81] J. J. van Griethuysen, A. Fedorov, C. Parmar, A. Hosny, N. Aucoin, V. Narayan, R. G. Beets-Tan, J.-C. Fillion-Robin, S. Pieper, and H. J. Aerts. Computational radiomics system to decode the radiographic phenotype. Cancer research, 77(21):e104–e107, 2017. BIBLIOGRAPHY 168

[82] F. H. P. van Velden, P. Cheebsumon, M. Yaqub, E. F. Smit, O. S. Hoekstra, A. A. Lam- mertsma, and R. Boellaard. Evaluation of a cumulative SUV-volume histogram method for parameterizing heterogeneous intratumoural FDG uptake in non-small cell lung cancer PET studies. European journal of nuclear medicine and molecular imaging, 38(9):1636–47, 2011.

[83] F. H. P. van Velden, G. M. Kramer, V. Frings, I. A. Nissen, E. R. Mulder, A. J. de Langen, O. S. Hoekstra, E. F. Smit, and R. Boellaard. Repeatability of Radiomic Features in Non-Small- Cell Lung Cancer [(18)F]FDG-PET/CT Studies: Impact of Reconstruction and Delineation. Molecular imaging and biology, 18(5):788–95, 2016.

[84] U. Vovk, F. Pernus, and B. Likar. A review of methods for correction of intensity inhomogen- eity in MRI. IEEE transactions on medical imaging, 26(3):405–21, 2007.

[85] R. L. Wahl, H. Jacene, Y. Kasamon, and M. A. Lodge. From RECIST to PERCIST: Evolving Considerations for PET response criteria in solid tumors. Journal of nuclear medicine, 50 Suppl 1(5):122S–50S, 2009.

[86] J. C. Waterton and L. Pylkkanen. Qualiﬁcation of imaging biomarkers for oncology drug development. European journal of cancer, 48(4):409–15, 2012.

[87] J. Yan, J. L. Chu-Shern, H. Y. Loi, L. K. Khor, A. K. Sinha, S. T. Quek, I. W. K. Tham, and D. Townsend. Impact of Image Reconstruction Settings on Texture Features in 18F-FDG PET. Journal of nuclear medicine, 56(11):1667–73, 2015.

[88] S. S. F. Yip and H. J. W. L. Aerts. Applications and limitations of radiomics. Physics in medicine and biology, 61(13):R150–66, 2016.

[89] C. Zhang and T. Chen. Eﬃcient feature extraction for 2D/3D objects in mesh representation. In Proceedings 2001 International Conference on Image Processing, volume 2, pages 935–938. IEEE, 2001.

[90] A. Zwanenburg, S. Leger, L. Agolli, K. Pilz, E. G. C. Troost, C. Richter, and S. L¨ock. Assessing robustness of radiomic features by image perturbation. Sci. Rep., 9(1):614, 2019.